speech recognition developers

Abstract: Test-time training (TTT) layers replace the linear fast-weight memory of modern recurrent architectures with a small neural network trained online by gradient descent. This nonlinear memory is more expressive than the linear memories of Gated DeltaNet and Mamba-2, but is difficult to parallelize: existing implementations such as TTT-MLP rely on sequential, block-wise weight updates that run several times slower than chunkwise-parallel linear recurrences. Moreover, their block-wise execution only approximates the token-by-token online recurrence, whereas linear recurrences admit an exact chunkwise-parallel evaluation. We introduce ParaTTT (PTTT), an algorithm for parallelizing arbitrarily deep nonlinear memories by iteratively refining estimates of the inner-loop gradient trajectory. Starting from a guess of the hidden-layer gradients, each refinement consists of three chunkwise-parallel passes: a forward pass that evaluates the resulting feature trajectory, an exact gated-delta-rule pass that computes the readout trajectory and its feature gradients, and a backward pass that propagates these gradients into a new estimate. The exact online gradient descent recurrence is a fixed point of the iteration, while decoding requires no refinement iterations. We prove geometric convergence for a two-layer variant under sufficient contraction conditions. In practice, two iterations suffice. With two refinement steps at 4K context, PTTT evaluates the inner-loop trajectory $32\times$ faster than TTT-MLP with the block approximation at matched error, while the full 340M model trains $4.9\times$ faster. Models trained at 340M and 1.3B parameters match or exceed Gated DeltaNet and Mamba-2 and remain competitive with same-size Transformers on language-modeling perplexity and zero-shot accuracy, with better recall than the linear recurrences.

Abstract: Exploiting temporal information in datasets with high temporal dependencies (such as spoken keyword recognition) by spiking models is still an open problem. Promising frameworks include Time-To-First coding (TTFS) and the inclusion of synaptic delays into learning. In TTFS, neurons are forced to spike at most once, forcing the network to found its learning on the spiking times of neurons rather than their spiking rate. Some models use the TTFS framework to directly compute the spiking time of Integrate-and-Fire (IF) or Leaky Integrate-and-Fire (LIF) neurons. However, most models built in the TTFS framework have not been tested against datasets containing temporal information. Little is known about the latency of these models on temporal datasets, where the latency is the time taken by the network to correctly classify a sample. In this work, we study the behaviour of Delay Neural Networks (DeNN) models, i.e. models that directly compute the spiking time of neurons by bypassing the membrane potential and thresholding mechanisms of spiking neurons. We show how to analytically control their latency, and empirically study the latency-accuracy tradeoff in a dynamic temporal dataset.

Abstract: The ability to represent temporal information is often hypothesized to underlie the efficiency and learning capabilities of biological neural networks. This motivates the development of spiking neural networks. However, it remains unclear how temporal information is actually represented in neural networks and how different forms of temporal representation affect performance and computational efficiency. We introduce the notion of the degree of temporal representation of a neural network, defined by whether its constituent functions (activation, learning, encoding, and decoding) explicitly manipulate time as a computational variable. This perspective yields a continuum ranging from non-temporal representations, where time does not explicitly appear as a computational variable in the network functions, to fully temporal representations, where computations explicitly operate on temporal variables. Using a fixed deep architecture, we compare different encoding, activation mechanisms and decoding with increasing temporal expressiveness: non-temporal weighted-sum activations used in usual artificial neural networks, non-temporal binary spiking activations used in spiking neural networks, and spike-time-based activations used in delay neural networks. Experiments on event-based datasets show that increasing the degree of temporal representation improves both performance and computational efficiency, allowing state-of-the-art performances on the Google Speech Commands dataset. We further analyze how temporal representations induce distinct forward and backward computational trajectories, leading to gradients that depend explicitly on the history of event times.

Abstract: Simple aggregation of independent responses ("wisdom of crowds") generally improves collective performance through statistical facilitation, but interaction among members can sometimes hinder group outcomes. However, under certain circumstances, coordination studies show that interacting groups can outperform their individual members and nominal groups, and interaction can also be more efficient than independence. We present a reservoir computing model to investigate how interaction can enhance joint performance in terms of accuracy and efficiency. Echo state networks were trained independently to identify Japanese speakers from vowel recordings and tested jointly with and without coupled feedback. Model pairs reproduced the dyadic advantage reported by Bahrami et al. (2010) with interaction via pooled-feedback, but not when independent responses were aggregated together. Model dynamics revealed that interaction was beneficial whereas aggregated outputs were prone to diverging from runaway feedback over time. In terms of efficiency, larger groups of smaller interacting models outperformed smaller groups of larger models while equating for total amount of group computation. Our results offer a formal framework for explaining when and why interaction enhances collective intelligence.

Abstract: Linear-time sequence models typically store long-context information in a recurrent state. A fixed-size state limits how much information can be stored, while hybrid architectures that add full attention for better recall incur quadratic sequence cost. Alternatively, sparse attention with continued pretraining (CPT) can extend the effective context length. Generalization beyond the trained length remains challenging, however, and CPT can be costly. We introduce FreeSolo, a native sparse attention mechanism with a short-range sliding window branch and a long-range branch that selects keypoint cells for direct token attention. Its linear-time variant, FS-L, learns to prune the prefix before query-dependent selection through short CPT. FS-B searches balanced trees with $O(N\log^2 N)$ algorithmic cost including radix-sort construction. FreeSolo demonstrates better zero-shot length extrapolation than the baselines without long-context fine-tuning.

Abstract: A fundamental challenge in cognitive science and neural computation is the binding problem. Compositional generalization can be achieved by factorizing representations into roles, the structural slots of a representation (e.g., serial position or syntactic function), and fillers, the contents occupying them (e.g., a phoneme or a word), which can be recombined freely to represent novel combinations. How neural systems implement role--filler binding, however, remains unknown. One of its simplest cases is phoneme sequencing, where phoneme identity (filler) must be bound to position (role) so that /gum/ and /mug/ are represented differently. A proposal from NLP, the `onion hypothesis' (Csordas et al., 2024), holds that the direction of the neural population state encodes token identity, while its magnitude encodes position. It was formulated for random token sequences; we extend it to natural language and test it in neural models and the human brain. We first evaluated it in an encoder-decoder trained to repeat English words, using representational geometry analyses and causal interventions. Magnitude alone proves incomplete: position is encoded by magnitude together with a systematic rotation of the state direction, a scheme we formalize as the `spiral hypothesis'. We then analyzed intracranial recordings from the auditory cortex of 20 participants listening to short sentences. Across superior temporal gyrus electrodes, the change in population response between consecutive phonemes decreases in magnitude as position advances, and its direction is aligned within consonants and within vowels but near orthogonal across the two classes, supporting the magnitude coding shared by both accounts. These results provide a geometric account of how neural systems bind phoneme identity to position, and show how neural-model hypotheses can yield testable predictions about human neural computation.

Abstract: Large language models can post-edit ASR by using linguistic context to compare N-best hypotheses and rewrite the transcript when needed, often improving WER without retraining the recognizer. However, prompt-only correction is brittle: it may over-edit when top-1 is correct and hallucinate when the reference is absent, and often needs very large models. We propose oracle-guided fine-tuning that turns N-best lists into supervision via three curated pools: corrections (reference in the list but not top-1) to teach selection, regularizers (top-1 correct) to discourage unnecessary rewrites, and hard negatives (reference absent) to suppress hallucination. Fine-tuning on this data provides more effective supervision than random subset fine-tuning for learning a hybrid select-or-rewrite policy, with the clearest gains from around 3B parameters on the proprietary dataset but less conclusive results on CommonVoice, a divergence suggesting that public read-speech text may overlap with LLM pretraining and mask genuine post-editing gains.

Abstract: Speech large language models (LLMs) are often built using pretrained acoustic encoders as an audio interface to an LLM decoder. However, LLMs scale poorly with input sequence length, while acoustic encoders tend to produce substantially longer sequences than typical textual inputs. This motivates learning to compress encoder outputs while preserving task-oriented information. To this end, we introduce ASSET, a reinforcement learning (RL)-based framework that learns to adaptively compress encoder outputs by first partitioning them into variable-duration segments and then pooling each segment into a continuous audio token. We use group-relative policy optimization (GRPO) to train a segmenter with task performance feedback and a token-rate penalty. In automatic speech recognition (ASR) experiments, ASSET achieves performance comparable to a transcript-assisted character-boundary oracle with approximately 25% fewer audio tokens on English LibriSpeech, while improving on a phone-boundary oracle at a similar token rate on Mandarin AISHELL. Exploratory experiments on non-ASR tasks further illustrate how learned segmentation changes when the framework is adapted to tasks beyond transcription.

Abstract: Power sampling improves large language model (LLM) reasoning at test time by sampling from a sharpened version of the base model's distribution. Its sampler, PowerMH, is slow because its Metropolis--Hastings (MH) updates are sequential: each accept/reject decision sets the state from which the next proposal is generated. Prior work reduces this cost by changing the sampler or the model (e.g., sequential Monte Carlo, approximate lookahead, distillation). We propose Predictive SubTree Prefetching (PreSTO), which accelerates PowerMH by prefetching future proposals and provably leaves its MH transition kernel unchanged. PreSTO views future MH outcomes as a tree of accept/reject decisions: in each model call, it generates the proposals of a \textit{prefetched subtree} of this tree in one batch, then follows the realized path with the original MH acceptance rule. We also observe that MH reaches only a sparse, nearly unary part of each prefetched subtree, we further design a \textit{predictive traversal rule} that chooses the subtree to maximize the predicted number of completed transitions. Across seven datasets and five base LLMs, PreSTO achieves median speedups of $1.71\times$, $1.91\times$, and $2.33\times$ on PowerMH and its follow-ups EntropyCut and MultiTryMH, respectively; on the first two, it makes a median of $3.9\times$ fewer model calls, each only $2.1\times$ longer. On LCB V6 with Qwen3.5-9B, the predictive traversal rule is a further $1.25\times$--$1.37\times$ faster than the default breadth-first traversal.

Abstract: Existing sequence models share a critical limitation: they lack an explicit, interpretable mechanism for content-aware temporal recall. LSTMs compress everything into a single hidden state, thereby losing information about specific past events. Transformers attend to all tokens indiscriminately at quadratic cost, while state-space models like Mamba compress history into a fixed-dimensional state without explicit content-addressable memory. To address this gap, we introduce a novel framework grounded in nonlinear time-delay systems, in which the effective delay is dynamically estimated based on the semantic similarity between current and historical inputs. Our approach features three key innovations: (i) a differentiable semantic lag estimator that computes content-aware soft attention weights over history, producing a weighted historical context and similarity score; (ii) a diagonal state space with closed-form cumulative sum solution, achieving O(1) per-step complexity with no sequential loops; and (iii) an interpretable fuzzy semantic modulation mechanism that learns context-dependent rules for balancing instantaneous processing versus historical recall, where rule centers correspond to prototypical semantic contexts. Experimental results on text classification and generation outperform state-of-the-art methods.

Abstract: We interpret Transformer networks as learned approximations to a hybrid probabilistic-inference procedure that combines max-product Loopy Belief Propagation (LBP) with Expectation--Maximization (EM). In this view, a Transformer performs inference in partially observed Markov Random Fields (MRFs) whose structure and parameters are not given a priori but are inferred from data. We show that self-attention implements a differentiable approximation to max-product LBP on an input-conditioned, dynamically structured MRF, with attention weights encoding learned structural affinities. We further establish an EM correspondence at the level of depth: each layer updates latent representations in an E-step-like forward pass, while the transition $θ^{(l)} \to θ^{(l+1)}$ across layers acts as an implicit M-step. Building on this interpretation, we introduce \emph{meta-models}, a design paradigm in which inputs generate instance-specific parameters, unifying mechanisms such as self-attention, mixture-of-experts, and gated units within a single framework. Empirical results support the proposed perspective.

Abstract: Phoneme-level models transcribe or generate speech as a sequence of phonemes, the smallest sound units that distinguish words. These models enable fine-grained pronunciation control and understanding, yet often fail on input that does not match any single training language, such as speech alternating between two languages, known as code-switching, or low-resource languages absent from training. We identify a systematic failure mode behind this, phonological interference: models assume the input is in a single language and impose its phonology, overriding local phoneme-level decisions that conflict with the assumed language. We measure interference by how often a model retains phonemes that one language has but the other lacks. On code-switched input, two phone recognizers (speech-to-phoneme models) and a phoneme-conditioned text-to-speech model lose 32% to 79% of these phonemes, but lose far fewer of the phonemes both languages share. On unseen languages, we find that phone recognizers impose the phonology of the training language they assign to the speech, and the more confident the assignment, the more they lose phonemes the unseen language has but the assigned language lacks. We probe the models' language estimate from their internal activations, and trace interference to a low dimensional subspace. On monolingual speech, steering this subspace toward another language makes the model lose the phonemes that only the original language uses and produce phonemes that only the target language has. We introduce windowed language estimation (WLE), an inference time repair that replaces the model's language estimate in this subspace with one computed from a short window around each position. On code-switched input, WLE removes 34% to 69% of the interference in all three models, and in the recognizers it leaves monolingual performance essentially unchanged.

Abstract: Spoken language identification (LID) aims to recognize the target language regardless of accent. In practice, however, LID models fine-tuned from self-supervised speech representations frequently confuse accents with languages, misclassifying non-native (L2) speech as the speaker's first language (L1). We show that non-native speech representations lie between native target-language and native L1 poles, causing systematic misclassification. To address this, we introduce a geometric projection that estimates an L1-bias direction solely from native speech and removes it before the frozen LID head. Across five MMS-LID models and non-native corpora, this projection substantially improves target language identification for L2-accented speech while preserving predictions for native speech. These results show that accent-induced L1 bias can be corrected directly within the representation space without L2 training data or model adaptation.

Abstract: Contextual biasing supplies an ASR system with a list of expected words at inference time, but existing methods rely on word boundaries that Japanese and Chinese do not provide. We present a boundary-free biasing decoder for frozen public CTC models, built on a character-level Aho-Corasick automaton, with no training and no second pass. Two evidence-based mechanisms replace the boundary: a depth-adaptive gate that sets how hard to push from match depth, and reading-space matching for when the audio is right but the characters are wrong. On Aishell-1 NE's hard R1 subset we reach 66.5% recall, above the trained CLAS baseline (64%), transferring to WenetSpeech and to a second architecture without retuning. We release the first open Japanese contextual-biasing benchmark, where biasing lifts rare-word recall by 25 points at precision above 97%, and still by 19 and 22 points against 1,000-word lists.

Abstract: Automatic Speech Recognition (ASR) systems often show uneven performance across demographic groups, and errors can be especially difficult to address for speakers belonging to multiple demographic groups. This work studies demographic-aware model merging for fair Speech-LLM-based ASR. Starting from a SLAM-ASR-based model, we fine-tune only the connector on demographic-specific subsets and merge the resulting subgroup-adapted connectors into a global model. We then identify critical cross-axis demographic pairs using subgroup WER and task-vector conflict, and apply intersection-specific correction vectors to the global merged model. Experiments on Fair-Speech show that global demographic merging improves overall WER over the base model, while intersection correction provides additional gains for several merging strategies. In particular, TIES with WER-based correction achieves the best overall WER, reducing it from 7.38\% to 5.13\%. Subgroup and disparity analyses further show that the proposed approach improves performance across demographic axes, while highlighting that lower average WER does not always imply reduced subgroup disparity.

Abstract: Contextual priors improve neural-to-language decoding by reshaping candidate scores. However, confidence is read from the same reshaped scores, so the errors a prior leaves behind can become more confident with no change in accuracy to reveal it. We study how a prior shapes confidence in speech retrieval on MEG-MASC and MOUS using local decoding scores, a contextual prior combined by additive shallow fusion, and the fused top-two margin as confidence. Among initially incorrect predictions, we find a confidence-ordering reversal: a larger margin makes a repair more likely when the correct candidate starts near the top of the local ranking, but less likely when it starts lower. On MEG-MASC, pooled correctness AUROC is 0.87, yet AUROC separating repairs from residual errors falls from 0.70 at initial ranks 2-3 to 0.39 at ranks 21-50. Errors starting beyond rank 20, inside the reversed region, make up 46.6% of all post-fusion errors. We propose a score-level account: a repair must first close the correct candidate's initial deficit, limiting its final margin, whereas a residual error can build a large margin between two incorrect candidates. A causal intervention that changes only the fusion weight moves the reversal to deeper ranks as predicted. Under a word-level LM prior, it keeps moving after accuracy gain peaks, so a weight chosen for accuracy does not settle confidence. Reading local and prior scores separately improves selective decoding: the decoder answers on 74.5% of windows instead of 56.7%, while 92% of output sets still contain the correct candidate. Confidence after contextual fusion should retain the local and contextual evidence behind each prediction, not just the fused scores. Project website: https://confidencereversal.github.io/; Code: https://github.com/AmadeusFake/NeuDecodingConfReversal

Abstract: Speech-LLMs often exhibit prompt overfitting, where models solely trained on automatic speech recognition (ASR) instruction fail to generalize to new instructions such as speech translation and continue to behave primarily as ASR system. We propose DirectSpeech2LLM, a simple end-to-end framework that preserves the instruction-following ability of the LLM on unseen tasks when conditioned on speech. It computes distance-based CTC loss over the frozen LLM embedding matrix and uses greedy CTC labels to derive geometrically and temporally aligned speech embeddings respectively as an input to the LLM. Trained solely on 960 hours of LibriSpeech ASR data, DirectSpeech2LLM outperforms the cascaded system on ASR (seen task) and generalizes zero-shot to speech translation and emotion recognition (two unseen tasks), closely matching the cascaded system upper bound on these two new instructions despite seeing neither during training. We also find that geometric alignment strength plays a smaller role than previously assumed, as our modified CTC loss is shown to provide sufficient implicit geometric grounding without requiring an explicit regression loss. Results are consistent across two LLM families and scale with both more training data and model capacity.

Abstract: We study sequential content sufficiency by investigating whether a representation retains the ordered target information available in its input. An information-theoretic decomposition separates input ambiguity, representation loss, and readout mismatch. We construct recoverable views where perfect agreement and joint isotropic Gaussianity coexist with zero target information, and establish limits imposed by deterministic canonical anchors. Token log-loss provides a one-sided information-loss bound; a fixed-penalty ridge analysis shows why rank alone cannot determine prediction risk. These results motivate CANOPE, a nonautoregressive framework with ordered latent canvases, canonical-token supervision, and geometric regularization. On 40,000 validation sequences, latent-agreement (PL0) and token-grounded (PL2) have nearly identical pooled ranks but reach 13.5% and 98.8% positional Recall@1, respectively, under strong natural corruption when the correct target length is provided. On 3,930 LJSpeech validation utterances, frozen PL2 with a trained MatchaTTS readout yields 21.54% word error rate (WER) on corrupted text, versus 99.22% for frozen PL0, while end-to-end MatchaTTS reaches 10.93%. These results show that geometric regularity alone does not guarantee recoverable sequential content or effective downstream access in the text settings studied here.

Abstract: Recurrent sequence models must decide how strongly to overwrite their memory at each token. Read as Bayesian filtering, this write is the gain of a Kalman update, set by uncertainty from two sources that pull it in opposite directions: volatility, how quickly the underlying associations change, and stochasticity, how noisy each observation of them is. First, we show that the update of gated delta-rule memories is the form this filter takes under isotropic uncertainty. Next, we introduce Voltic, a recurrent memory that keeps the covariance anisotropic and makes both noise variances input-dependent, so the write is vector-valued and carries uncertainty accumulated over the sequence. A dense covariance would have to be propagated token by token, ruling out the parallel training these models depend on. We therefore give two assumed-density approximations, diagonal and quasi-diagonal, both of which leave the memory update in delta-rule form and reuse its chunked kernels. On controlled recall tasks in which associations change and observations are corrupted, Voltic leads all baselines. On the task combining volatility and stochasticity, its margin over the strongest baseline is larger at both extrapolation sizes than at the training sizes. In 45M-parameter language models it leads an eight-task reasoning average and achieves higher retrieval accuracy beyond the training context length than gated baselines, at throughput close to those baselines. Deriving the write from an uncertainty recursion therefore makes memory more responsive to change.

Abstract: Frozen speech foundation models (SFMs) make downstream adaptation efficient: the backbone can stay fixed while a task learns layer fusion and a lightweight classifier. Full adversarial fine-tuning is a standard route to robustness, but generating adversarial examples and updating the backbone for every task sacrifices that efficiency. We ask whether robustness can instead be learned before future tasks are known. For a frozen backbone and linear classifier, robustness can be understood through the interaction between representation stability and decision-boundary margin. This leads directly to our design: we stabilize representations across the hidden layers, rather than only the final layer, while preserving clean representations; after clean adaptation selects the layer mixture, we keep it fixed and enlarge only the classifier margin, without downstream adversarial examples. We evaluate Wav2Vec2, HuBERT, and WavLM Large on four tasks under adaptive 30 dB attacks. Across 12 backbone-task pairs, hierarchical robustification improves robust accuracy by 46.4 pp, while margin refinement adds 4.0 pp for 1.1 pp of clean accuracy. Code and configurations are available at https://github.com/arefmousavi/hierarchical-robust-sfm.

Abstract: Knowledge distillation (KD) improves low-resource acoustic learning by enriching one-hot supervision with the softened predictive distribution of a fixed teacher network. However, a teacher trained with limited or imbalanced annotations may produce a biased distribution whose components are not uniformly reliable. Although this distribution can still encode useful knowledge, direct full-distribution matching may also transfer teacher-induced biases, thereby distorting the student's decision boundary and degrading its generalization performance. To address this limitation, we propose Boundary-Anchored Mass-Partitioned Distillation (BA-MPD), a logit-based distillation objective composed of Boundary-Anchored Correction (BAC) and Mass-Partitioned Distillation (MPD). BAC addresses missing ground-truth labels in the set of the teacher's top predictions by swapping the true label for the lowest-ranked entry of the set, thus keeping the mass and uncertainty of the set unchanged. MPD then distills this corrected distribution through separate losses that enforce relational consistency within the set, balance the mass between high- and low-confidence groups, and weight lower-confidence dependencies. Ultimately, BAC and MPD together suppress harmful ranking errors and noisy low-confidence details, while retaining all useful teacher information. Experiments on two acoustic benchmarks under multiple label budgets show that BA-MPD consistently improves over supervised-learning baselines and vanilla KD while remaining competitive with strong logit-based KD baselines. Cross-budget results further show that BA-MPD remains effective when the teacher and student models use mismatched label budgets, demonstrating its ability to exploit imperfect teachers across supervision gaps. Implementation available at https://github.com/ShuanglinLi/BA-MPD.

Abstract: In a Transformer, token mixing is the step that lets each token draw information from other tokens, and it dominates the cost of encoding long sequences. Self-attention does this mixing very well: every token weighs every other token by content, which gives strong contextual modeling. That all-pairs comparison is also why its cost grows quadratically with sequence length. We introduce WNet, a Transformer encoder that replaces self-attention with token mixing based on the discrete wavelet transform (DWT). Three attention-free mixers recombine the scales: by linear fusion, by learned gating, or by letting each token choose its scales. A hybrid adds self-attention in the last layer only. A receptive-field analysis shows that wavelet mixers built from two-tap filters, such as Haar, never relate tokens outside fixed blocks, however deep the network, even when the filters are learned. Longer filters reach the whole sequence within two layers. We pre-train every model with masked language modeling on a fixed-token subset of C4 and fine-tune on GLUE, using one controlled setup with size-matched BERT and FNet baselines and a control that cannot mix tokens. The token-gated mixer trains as fast as attention at 256 tokens and 2.7 times faster at 4,096.

Abstract: Source-Free Universal Domain Adaptation (SF-UniDA) extends Universal Domain Adaptation by removing access to source data at adaptation time while still handling label-set mismatches between domains. Despite growing interest in this setting for image data, no benchmark exists for time series, which are more challenging. We present the first SF-UniDA benchmark on time series. In addition, we provide the first study of pretrained foundation models as feature extractors for time series domain adaptation. In this context, we identify a critical and previously underexplored limitation of all existing SF-UniDA methods: the inference threshold for unknown-sample rejection is highly sensitive. We address this by proposing a plug-in auto-thresholding module that can be integrated into any SF-UniDA method. Experiments on three well-known time series datasets confirm the suitability of this module. They also highlight that foundation models do not systematically outperform classical backbones and that SF-UniDA tailored for time series is yet to be developed.

Abstract: The attention operation is naively position invariant. However, positional information is fundamental to natural language, and therefore a variety of explicit position encodings have been developed in transformer-based models, such as rotary position encoding (RoPE). Although explicit position encodings have long been assumed to be required, recent methods that interleave local mixing layers, such as sliding window attention (SWA) and gated linear attention, while not encoding position (NoPE) in global attention layers has recently been shown to be successful at scale. How and why this approach works is not well-understood. In this paper, we develop an explanation of how hybrid models of this sort can implicitly encode position at global NoPE layers. Supported by both theoretical and empirical evidence, our central argument is that SWA and gated linear attention induce a recency bias in the residual stream that propagates to, and is selected by, the global attention logits. Moreover, in contrast to the implicit position encodings found in models with only global NoPE attention, in which positional information arises solely from the causal mask, the recency bias in hybrid models can be maintained across long sequences. In addition to deepening our understanding of how hybrid models encode position, these findings may provide insights for how to encode position in a way that can extrapolate to longer sequence lengths indefinitely.

Abstract: Transcribing domain-specific entities and rare proper nouns remains a major challenge in automatic speech recognition (ASR). In this paper, we propose BaLEEN (Biasing with Latent Encoded Entities), a lightweight, hypernetwork-based framework for dynamic contextual adaptation without fine-tuning the underlying ASR model. BaLEEN encodes variable-length contextual keywords using a pretrained language model, compresses them into a fixed sequence of latent vectors via a Perceiver bottleneck, and injects context-dependent bias vectors directly into the intermediate encoder representations of the ASR model. Because both the language model and the backbone ASR model remain entirely frozen during training, BaLEEN operates as a plug-and-play adapter that incurs zero computational overhead at inference time when context biases are precomputed. We evaluate our method on a CTC-based ASR model using a Wikipedia-derived corpus with annotated named entities and synthetic speech. Experimental results demonstrate that BaLEEN reduces keyword miss rate by 8.7% on the test set relative to the unbiased baseline while simultaneously improving overall word error rate by 21% and character error rate by 28%.

Abstract: Recurrent neural networks (RNNs) compress the historical context into a memory state of fixed size, thus allowing for constant-time inference. The memory state size is a crucial factor in their performance, as exemplified by the strong performance and resurgence of linear attention, which extends the vector-valued hidden states of ordinary RNNs to matrix-valued hidden states. Crucially, linear attention does so in a parameter-efficient way, in particular by using an outer product of the key and value vectors to write to the matrix-valued hidden state. We generalize this construction and propose triadic linear attention, which writes the triadic outer product of a key, a second key, and a value, into a third-order (i.e., 3D) tensor state, and reads from it by contracting both key axes with two queries. An $E$-dimensional second key thus yields an $E$-fold increase in state size while adding only two projections. Triadic linear attention is compatible with data-dependent forgetting, the delta rule, and chunkwise-parallel training. Applied to Gated DeltaNet and scalar-gated linear attention, triadic linear attention substantially improves long-context language modeling and recall, outperforming alternatives that enlarge the state.

Abstract: Linear RNNs offer linear-time sequence processing and constant-memory decoding, but their fixed-size recurrent states must accommodate all past key--value associations. Existing forgetting mechanisms and Delta Rule updates reduce interference by selectively clearing or correcting the state, yet earlier associations can still become difficult to retrieve. We introduce CyFA (Cyclic Flow Attention), a Linear RNN with relative-time-partitioned memory. At each step, a learned clock controls the cyclic transport applied jointly to the key and value states before the current key--value pair enters the age-zero slot, thereby organizing stored associations across relative-time slots. We further derive an exact change to absolute-clock coordinates that expresses CyFA as two scalar-decay linear attention recurrences and enables efficient chunk-wise training. Across 400M--1.4B pretraining experiments with matched recurrent-state sizes, CyFA improves recall-intensive performance while maintaining competitive language modeling and high computational efficiency. At 400M, CyFA outperforms KDA on FDA (42.60 vs. 26.07) while requiring only 46.7% and 48.3% of KDA's forward and backward core-operator execution times, respectively. Our code is publicly available at \href{https://github.com/Chyxx/CyclicFlowAttention}{this https URL}.

Abstract: Looped transformers have demonstrated promising parameter efficiency by reusing layers for latent computation. Prior studies compare looped and non-looped models at matched parameters or per-token FLOPs. However, to the best of our knowledge, whether looping improves test-time scaling as outputs grow longer remains underexplored. Through post-training looped transformers, we study the accuracy-compute slope, measured as the accuracy gain per doubling of test-time decoding FLOPs. We find that existing looped transformers often yield steeper slopes than their non-looped baseline, yet underperform it at matched compute. While fixed-depth looping spends extra iterations on every token, our analysis shows that many tokens do not benefit from extra iterations. We therefore propose TaH2, which enables the model to focus extra iterations on the tokens that benefit from looping. It jointly post-trains the backbone and an iteration decider through lookahead depth supervision, which uses online labels indicating whether further iteration improves the prediction. TaH2 improves both the efficiency and attainable accuracy of test-time scaling. On challenging AIME benchmarks, TaH2 improves the accuracy-compute slope by 53% (2.74 vs. 1.79) over the non-looped baseline, exceeding the baseline's peak accuracy by about 3.4 points at matched test-time compute. As the maximum iteration depth increases, existing looped models largely plateau, while TaH2's gain over the non-looped baseline continues to grow from +2.8 points at depth 2 to +3.9 points at depth 8. Our code is available at https://github.com/thu-nics/TaH.

Abstract: With the recent advancements of neural compressors, explicitly incorporating perception constraints into the design of compression schemes has gained significant attention. Traditionally, these perception constraints ensure that the distribution of the reconstruction does not significantly deviate from the distribution of the source, thus attesting to the perceptual quality of the reconstruction. In this work, we uncover several perception constraints that are naturally present in task-aware compression. In particular, we consider a problem where the primary task is reconstruction and the secondary task is classification (i.e., a statistical test). We study this problem at varying levels of domain information available to us and discuss how to utilize the naturally emerging perception constraints to design rate-minimal compression schemes that also maximize the utility of our secondary task. We show that in this setting, if the decision boundaries of the classifier are ill-defined (mismatch) for our source distribution, then matching onto a target distribution enhances our classification accuracy.

Abstract: Sign language translation and generation share the goal of bidirectional alignment between text and sign representations. However, existing approaches either treat them as isolated tasks or are only verified on limited datasets, limiting effective modeling between modalities. In this paper, we propose SignFLIP, a unified LLM-centered framework for translation and generation. To enable bidirectional mapping between text and sign, SignFLIP adopts a symmetric architecture together with a stage-wise training strategy built on large-scale data. The shared sign--text representation is progressively refined: pre-alignment facilitates subsequent SLT, while the SLT-adapted representation further benefits SLG. Extensive experiments on multiple benchmarks show that SignFLIP shows competitive performance compared with task-specific models on both translation and generation tasks, as well as strong transferability to sign language recognition.

Abstract: We present Cardamom, a micro-dialectal Arabic speech dataset designed to support fine-grained evaluation and adaptation of automatic speech recognition (ASR) systems. Community-curated by native speakers familiar with the represented varieties, Cardamom contains approximately 40 hours of transcribed YouTube speech spanning 21 micro-dialects across Egypt, Jordan, Lebanon, Mauritania, Palestine, and Saudi Arabia. Each segment is annotated with one or more operational micro-dialect labels, code-switching information, and utterance-level perceived gender, enabling analysis of sub-country variation that is obscured by conventional country-level labels. We describe the collection and annotation process, motivate the micro-dialect inventory linguistically, and benchmark four multilingual ASR systems in zero-shot and adapted settings. The strongest zero-shot system obtains 43.47% aggregate WER, with particularly high error rates on Mauritanian and Lebanese varieties; adaptation on Cardamom reduces its WER to 35.21%. Audio-based identification experiments further show that the annotations provide a learnable prediction target, with a dedicated classifier reaching 85.57% accuracy on 21-way micro-dialect identification. Cardamom provides a resource for studying localized dialectal variation and developing Arabic speech systems with broader regional coverage.

Abstract: Perturbation-based XAI methods for audio classifiers often estimate feature importance by masking spectrogram regions and crediting output changes to the retained signal. Yet they typically assume the fill (the mask replacement) is negligible. We propose logit-space trajectory decomposition to examine this assumption. An on-axis component captures output along a line connecting the fully filled (occluded) spectrogram to the fully retained original; an off-axis component captures perpendicular displacement. We evaluate across three fills, three audio classifiers, and 1,100 AudioSet clips. We demonstrate that no fill is acoustically neutral under full occlusion: Zero fill activates silence and Gaussian Noise fill activates broadband noise. Under partial masking, 41-77% of output displacement is off-axis, with the direction of the residual stable across mask retention fraction and specific to each model-fill combination. Attribution methods are unevenly exposed to off-axis displacement through their sampling and weighting strategies, revealing apparatus-dependence. Where the off-axis residual is stable and low-dimensional, its structure affords mitigation.

Abstract: A byte-level BPE vocabulary learns each inflected form of a Nepali word as a separate string, so a noun stem is spelled differently in each of its case-marked forms. We test whether splitting words into stem and affixes before BPE helps, with the corpus, vocabulary size, model and number of training steps held fixed. Our pre-tokenizer, Papaya, uses a finite-state transducer built from a published grammar of Nepali, falls back to regular expressions, and leaves the BPE trainer unchanged. On 607 words annotated by seven native speakers its segmenter reaches 0.96 boundary F1, and the resulting tokens keep stems intact far more often than plain BPE does. In a 17M-parameter language model it lowers bits per byte by about 1% at equal training steps; most of the larger gain seen at equal epochs comes from the extra steps that longer token sequences buy, and an unsupervised Morfessor segmentation gives the same improvement. Downstream the effect is small: NER improves only on entities that contain words unseen in training, POS tagging and news classification do not change, and published Nepali tokenizers perform about as well. We release the annotated boundary set, a 556-affix dataset and the code.

Abstract: Whisper is still what runs in production: one permissively licensed checkpoint, 99 languages, no per-language tuning. But it writes sentences nobody said. On 42 clips of pure room tone, whisper-large-v3 emits words on 61.9% of them and emits something on 100%. The usual response is to distil a student from Whisper pseudo-labels, filtering the hallucinations out of the corpus first, which deletes the evidence while leaving the behaviour in place: the student is fitted to the subset where the teacher was right, and inherits a failure mode absent from its own training data. The second response, adding non-speech audio so the model learns to stay quiet, already ships and works, but teaches suppression without discrimination. The checkpoint best on every non-speech arm here is also the one that recovers the fewest genuinely spoken phrases and deletes 58% of repeated speech. The teacher has to be fixed, with both halves of the signal. Our benchmark scores both at once: 11,852 clips over eight arms, 6,267 synthetic positives, 8,296 clips of real audio that made a production model fail, and FLEURS in 58 languages. We collect 40,891 hallucination phrases in 100 languages, choose a text-to-speech system by measurement, and synthesise those phrases as positives, so the model meets the same text both as something to suppress and as something to transcribe. Across 33 matched pairs of fine-tunes differing only by those positives, adding them lowers word emission on real voice-free audio in 31 pairs and raises phrase recovery in 32; among the 30 pairs that hold English accuracy it is 30 out of 30. The best checkpoint takes hallucination on silence from 61.9% to 2.4% and words over real voice-free audio from 99.9% to 47.8%, while raising phrase recovery from 69.8% to 82.7%. The benchmark, lexicon and synthetic corpus are released at https://huggingface.co/datasets/Scicom-intl/Whisper-Hallucination

Abstract: Many real-world time series data exhibit structural properties at multiple scales, from short, recurring units to complex sequences composed of these units. Unsupervised discovery of both these components and structure enables the design of intelligent systems that help interpret temporal data, thereby limiting the amount of costly human annotations required. Existing modeling approaches typically overlook the hierarchical structure inherent to many time series, treating recurring patterns at different temporal scales as independent structures. Moreover, most assume access to the complete data sequence and treat discovery as a static process, limiting their ability to evolve as new observations arrive. We introduce Rondo, an unsupervised approach for modeling recurring hierarchical structure in continuous temporal streams. By explicitly constructing vocabularies of reusable units and their recurring compositions, Rondo captures structure shared across complex temporal patterns while refining and expanding its discoveries as the stream evolves. Evaluations on temporal sequences spanning diverse domains and data modalities show that Rondo outperforms existing unsupervised recurrence-discovery baselines, with particularly pronounced advantages in limited-data and continual-stream settings. These capabilities provide a stronger foundation for recurring-pattern discovery, scalable behavior understanding, and adaptive intelligent systems operating on long, unlabeled temporal streams.

Abstract: Audiovisual Speech Recognition (AVSR) is a multimodal approach to speech recognition that incorporates visual information from lip movements to enhance model performance. Despite its advantages, its development remains constrained by the limited availability of labeled audiovisual (AV) datasets. This work explores the use of synthetic visual data as a solution, using an audio-driven talking-head pipeline to generate lip-synchronized visual content from existing audio data. We evaluate the effectiveness of synthetic visual data both as an augmentation strategy and as a standalone training resource, applying our approach to Spanish and Catalan. Our results show that augmenting real AV data with synthetic samples yields relative Word Error Rate (WER) reductions of up to 16.2%, demonstrating the potential of this approach. Moreover, we demonstrate that synthetic data alone can serve as a baseline for AVSR training in languages lacking AV datasets. These findings provide evidence that synthetic visual data can serve as a scalable solution to AVSR data scarcity, enabling broader language coverage.

Abstract: Full-duplex speech language models continuously accumulate acoustic key-value (KV) states, making long-running interactions memory-intensive. During listening, the model can finish processing an audio unit before the next arrives; we term the remaining interval listening-time slack. We propose acoustic-to-text KV compression, which introduces a transcription side channel to convert incoming speech into compact textual memory within this interval. When the cache exceeds a target budget during inference, older acoustic states are evicted while transcripts and recent acoustic context remain. We train the side channel with LoRA using cross-entropy on transcription segments. To preserve listening and speaking behavior, we apply knowledge distillation to the original model's token-level output distributions at native prediction positions. On ten-minute LongSpeech sessions, our MiniCPM-o 4.5 implementation reduces peak streaming KV-cache size by 64.6% compared with the same model without eviction. The proposed method also improves transcription, temporal question answering, and summarization over the baseline. Full-Duplex-Bench evaluations further show comparable pause-handling, turn-taking, and interruption performance.

Abstract: Accurate and efficient pronunciation transcription is essential for preparing text-to-speech training data at scale. Existing approaches have different limitations: grapheme-to-pronunciation (G2P) and speech-to-pronunciation (S2P) methods each capture only partial information, using only text or only speech, while speech-and-text-to-pronunciation (ST2P) methods use both but require costly pronunciation-annotated data. To address this problem, we propose a training-free ST2P pipeline that integrates both lexical and acoustic information at inference time. Lexical resources and G2P tools generate text-constrained candidates, and a left-to-right greedy search selects the best one using whole-sequence negative log-likelihoods from frozen pretrained S2P models. On three Japanese corpora, our method reduces Character Error Rate (CER) from 0.60--1.40\% (text-only baseline) to 0.04--0.17\% with reference transcripts, and 0.64--1.58\% with ASR transcripts. It outperforms all baselines, including a trained ST2P model and commercial multimodal LLMs. Our greedy search method is 3--3.5$\times$ faster than beam search at similar CER, and the cascade is 2$\times$ faster than direct decoding ensuring the efficiency and accuracy. In Spanish, French, and preliminary English, it also surpasses four open multimodal LLMs and the best traditional methods.

Abstract: Automatic speech recognition is typically trained assuming that the reference transcript is the only valid labeling of an utterance, yet even nominally verbatim transcripts contain localized differences in pronunciation, spelling, or lexical realization that the acoustics do not uniquely determine. Omni-temporal Classification (OTC) tolerates such noise by adding wildcard paths to the connectionist temporal classification (CTC) alignment graph, but its word-level arcs are too coarse, since bypassing one unsupported token discards supervision for the whole word. We move wildcard arcs to token granularity so unsupported tokens can be bypassed while the rest of the word stays supervised, and we combine token- and word-level arcs as complementary escape paths. Across 19 languages and three corpora, token-level OTC improves over CTC on all 25 tasks. We also replace epoch-indexed relaxation of the wildcard weights with a predictive-entropy-indexed schedule, which performs comparably while reducing dependence on training length. Combining this schedule with the hybrid graph gives the lowest mean word error rate (WER) on every corpus and a 9.45% average relative WER reduction over CTC. Independent validator transcriptions show that token-level models place significantly more wildcard-bypass probability than CTC on disputed characters, indicating that token-level tolerance targets localized transcript ambiguity.

Abstract: Vietnamese speech research is constrained by resources that isolate automatic speech recognition from speaker, dialect, code-switching, and deepfake analysis. We introduce VietPrism, an open, multi-domain corpus that brings these dimensions together at scale: 993.4 hours and 403,941 bona fide utterances from 1,262 verified speakers across 8,388 real-world videos. To our knowledge, it is the first large-scale Vietnamese corpus to jointly provide transcripts, consistent speaker identities, five dialect groups, and naturally occurring Vietnamese--English code-switching, which constitutes nearly half of the corpus by duration. We further create over 3.1K hours of spoof speech with four open-source and commercial synthesis systems. Every spoof is conditioned on a verified speaker reference and paired with a transcript- and speaker-matched bona fide utterance, enabling unique controlled evaluation with reduced lexical and identity confounds. Zero-shot evaluation of five pretrained multilingual detectors reveals striking brittleness: EER greatly varies across detector--generator pairings, while recent multilingual detector DFA-1B degrades from 16.3% to 33.6% as speaker similarity increases. Dialect-stratified results expose further model-dependent disparities. By unifying natural linguistic diversity with controlled spoof generation, VietPrism provides a challenging foundation for Vietnamese speech modeling and trustworthy audio-deepfake detection.

Abstract: Contextual biasing improves rare-word recognition in speech large language models (SpeechLLMs), but efficiently exploiting large bias lists remains challenging. We propose PTC-Bias, a two-stage framework based on phoneme-level temporal competition. At the prefill stage, PTC Retrieval performs frame-synchronous phoneme decoding and temporal competition among candidate pronunciations, producing a compact bias-word shortlist and corresponding speech intervals. After SpeechLLM decoding, PTC Correction conducts a second local competition between the retrieved candidates and mismatched transcript spans within these intervals. Selective correction reduces near-homophone and word-segmentation errors while preserving correct transcriptions. Both stages share the same phoneme posteriors and require no additional SpeechLLM forward pass. Experiments on LibriSpeech show consistent gains across two SpeechLLMs and bias lists of up to 2000 words. With Prompt-SLAM-ASR-7B and 2000 bias words, PTC-Bias reduces B-WER by 23.4%/23.9% relative to CTC-Filter on test-clean/test-other, while keeping U-WER nearly unchanged.

Abstract: Language modeling using Transformers has become commonplace despite their fixed computational depth and quadratic runtime with respect to input tokens. Recurrent models on the other hand offer linear depth but no parallel execution. In this work, we extend balanced-tree recursive operators from sequence encoding to autoregressive prediction, enabling all prefix representations to be computed with logarithmic depth and linear runtime. Our experiments provide an initial characterization of this model class, demonstrating robust length extrapolation and performance approaching that of ALiBi-based Transformers, highlighting its potential as an alternative architecture for language modeling.

Abstract: Conventional Japanese automatic speech recognition (ASR) is supervised by an orthographic transcript, although the same written form can correspond to different lexical readings realized in speech. Such utterances receive an identical target, so their reading distinction is absent from the supervision interface and cannot be recovered reliably by post-hoc text-only grapheme-to-phoneme conversion. We present Ruby-ASR, which refines the conventional target into a span-bound orthographic--lexical-reading sequence. Unlike separate full-sentence orthographic and phonological outputs, the ruby representation locally binds each written span to its realized reading and permits deterministic recovery of both views. We instantiate the target under subtitle-style and verbatim-style transcription conventions using a Qwen3-ASR backbone; a mora-level CTC objective provides auxiliary monotonic reading supervision. The experimental results across five Japanese benchmarks show that refining the recognition target can improve lexical-reading recovery without sacrificing readable orthographic transcription. We release the checkpoints and inference code.

Abstract: Angular-margin speaker encoders are widely used in voice conversion, yet the geometry of their classifier prototypes remains poorly understood. We analyze ECAPA-TDNN classifier prototypes as points on the unit hypersphere and characterize their organization using rotation-invariant angular statistics together with global and local effective dimensionality measures. Our analysis shows that standard training can induce angular concentration and a substantial reduction in effective dimensionality. To address this, we investigate two geometric regularization strategies (hinged Riesz log-energy and effective-dimension maximization) applied to classifier prototypes to encourage more uniform hyperspherical coverage. The resulting prototype sets exhibit higher effective dimensionality and improved isotropy, with configuration-dependent effects on speaker-recognition performance. When the corresponding ECAPA-TDNN models are used as speaker encoders for Fast-VGAN, the regularized systems also exhibit improved robustness in zero-shot voice conversion, particularly for previously unseen speakers.

Abstract: We reconsider the theory of probabilistic formal languages generated by n-gram models and by probabilistic context-free grammars (PCFGs). The expected hierarchy of probabilistic grammars is established by proving that every probabilistic language generated by an n-gram model is also generated by some PCFG, while some probabilistic languages generated by PCFGs cannot be generated by any $n$-gram model. We introduce the notion of fully connected PCFGs, namely PCFGs in Chomsky normal form where every production rule only involving non-terminals has non-zero probability. Our main result shows that any probabilistic language generated by an $n$-gram model differs from any probabilistic language generated by a fully connected PCFG. Therefore, the class of probabilistic languages generated by $n$-gram models is not a subset of the class generated by fully connected PCFGs.

Abstract: Speech-text space alignment is a multimodal representation learning method consisting to map different speech and text into a shared representation space, leading to enrichment of the representation of each modality. Proposed architectures, such as SAMU-XLSR, typically follow a student/teacher framework, with the goal of fine-tuning an audio encoder to produce representations that closely match those of the text. In this way a speech representation is semantically enriched. However, such systems generally require large amounts of training data and considerable computational resource, making them difficult to apply to low resources languages under frugal constraints. The present work proposes a data-efficient space alignment method based on Heterogeneous Graph Neural Networks and link prediction. The core idea is to leverage message passing to explicitly transfer information from the text modality to the speech modality, thereby reducing the need for large training datasets and intrinsically enriching the acoustic representations, all in a more interpretable manner. Although thoroughly explored for high-resource languages, word-level tasks in speech remain relevant for certain low-resource languages. Therefore, we conducted experiments on speech-text alignment at the word level using the TIMIT (English) dataset and Yemba (a Cameroonian language). Our approach yields results comparable to those of SAMU-XLSR, a state-of-the-art method, and even surpasses it for the Yemba language in the task of word retrieval, while using far fewer resources, demonstrating its power, frugality, and efficiency.

Abstract: Word Error Rate (WER), the most commonly used metric for Automatic Speech Recognition (ASR), treats every lexical deviation from the reference as equally costly, regardless of whether it changes meaning. This raises the question: does WER actually track how humans judge ASR transcript quality? We introduce HATS-en, an English dataset for human-centered ASR evaluation. Using this dataset, we benchmark lexical metrics against several configurations of BERTScore and SemDist, varying the language model, layer, and pooling strategy. We find that WER agrees least with human judgment among all metrics tested, that the best-performing SemDist configurations achieve the highest overall agreement, ahead of CER and BERTScore, and that no single model is best across settings. CER, despite its simplicity and low cost, remains remarkably close to these best configurations. In line with prior recommendations, our results support shifting ASR evaluation toward CER both for English and for morphosyllabic writing systems as it is a more interpretable and low-cost metric for what evaluation should actually capture, and using SemDist as a complementary evaluation.

Abstract: Conventional tokenizers represent text as characters or statistically derived subwords, overlooking the internal phonological structure of syllables and often requiring large vocabularies. We introduce \textbf{Phonemic Tokenizer}, a linguistically motivated tokenizer for Vietnamese and Chinese that converts each syllable into IPA and factorizes it into three phonological components: onset, rime, and tone. The three components jointly occupy one contextual position, preserving syllable-level sequence length while enabling representation sharing across phonologically related syllables. Non-phonological and unsupported units are handled through character-level fallback. This deterministic design requires no corpus-dependent vocabulary learning and yields vocabularies of only 112 entries for Chinese and 256 for Vietnamese. Intrinsic evaluation shows that the tokenizer achieves substantially higher Rényi efficiency in both languages, represents every entry in a standard Vietnamese syllable dictionary with a Fertility of exactly one, and generally produces shorter Vietnamese sequences than existing pretrained tokenizers. We further instantiate the tokenizer in \textbf{PhonemicBERT}, which combines factorized component embeddings and reconstructs complete masked syllables using three prediction heads. Under a controlled Chinese pretraining setup, PhonemicBERT-Zh is competitive with or outperforms character, subword, and SubChar alternatives across diverse language-understanding tasks. PhonemicBERT-Vi also achieves competitive or superior results to established Vietnamese and multilingual pretrained models. These results establish phonemic factorization as a compact, efficient, and interpretable alternative to atomic and statistically segmented text representations.

Abstract: Modern automatic speech recognition (ASR) systems trained on extremely large datasets can produce transcripts with numbers written in Arabic numerals. This creates a need for fair comparison with models that output verbatim texts and proper processing of reference transcripts. Popular approaches often reduce text normalization to lowercase and remove punctuation, with no additional normalization applied to languages other than English. In this work, we analyze the impact of normalization of numerical expressions in the evaluation of ASR systems in various languages, using Polish as an example of a highly inflective language. We perform experiments on VoxPopuli and The Polish Parliamentary speech datasets and estimate word error rate (WER) differences for different text normalization approaches. We show that the difference due to the lack of number normalization in WER may be substantial - more than 2 percentage points, and often higher than the differences between systems in popular multilingual benchmarks.

Abstract: Grapheme-to-phoneme (G2P) conversion turns raw text into its phonemic form and is an essential part of both text-to-speech (TTS) and automatic speech recognition (ASR) systems. It is required to be fast, stable and context-aware. For unsegmented languages such as Japanese, G2P additionally couples word segmentation with highly context-dependent polyphone disambiguation, and the scarcity of accurately annotated data remains a bottleneck. In this paper, we present a context-aware neural G2P method that scores paths of a discriminative conditional random field (CRF) over a word lattice constructed from dictionaries. To tackle data scarcity, we utilize large language models (LLMs) to generate more than 2 million sentences. Experimental results demonstrate that our method strongly outperforms conventional morphological analyzer-based methods and neural sequence models. On the Joyo-Kanji-Yomi benchmark, our method reaches 99.62% target word reading accuracy, 0.32% target word phoneme error rate (PER) and 0.14% sentence PER.

Abstract: Streaming automatic speech recognition (ASR) must be judged jointly on what it transcribes and on how quickly it commits each word. Delayed streams modeling (DSM) has become the dominant paradigm for streaming large audio-language models, exposing a structural delay $τ$ that bounds the decoder's lookahead. We show that $τ$ is a poor proxy for user-perceived latency, and that the alignment-based supervision of DSM leaves latency on the table: the same forced-aligned transcript is used at every $τ$, forcing the model to withhold words it could already commit. We introduce AWED, a word-level emission-delay metric defined relative to the acoustic end of each word, and post-train a DSM recognizer with GRPO under a reward that scores transcription accuracy and measured delay jointly. Trained at a single operating point ($τ=6$ frames), our model dominates both its supervised fine-tuning initialization and the Voxtral Realtime backbone across all evaluated lookahead budgets: it cuts WER by 30.8\% relative at an 80\,ms structural delay, and by 5.7\% relative at 480\,ms while lowering median AWED from 1.17\,s to 1.04\,s. Latency-rewarded post-training thus advances the accuracy--latency Pareto frontier of streaming ASR without architectural change.

Abstract: In Taiwanese Hokkien automatic speech recognition (ASR), prior studies often treat tone sandhi as a major challenge under the assumption that models fail to process implicit phonological variations. However, our experiments on Taiwanese Hokkien reveal that speech foundation models actually handle tone sandhi variations effectively, and the real performance bottleneck stems from a localized confusion between these variations and retained citation tones. To address this, we propose T-SANDHI to explicitly decouple surface acoustics from underlying lexical intent on top of a frozen Whisper backbone. Using a lexicon-guided multi-task learning structure driven by text-derived pseudo labels, our lightweight hybrid injection module integrates independent citation and sandhi phonetic streams via dynamic gating. Extensive evaluation on the TAT-MOE corpus and two blind test sets demonstrates that this explicit disentanglement effectively resolves tonal mapping confusion, outperforming baselines with strict parameter efficiency.

Abstract: Autoregressive generation of interleaved text and acoustic tokens is a common approach to spoken-response generation in speech large language models. Although this design enables streaming generation with explicit textual guidance, generated acoustic tokens become part of the context for subsequent text predictions. Given identical speech inputs, we observe markedly lower answer accuracy for the internal text generated in speech-to-text-and-speech (S2TS) mode than for speech-to-text (S2T) responses. We term this discrepancy the \emph{output-mode gap} (OMG). To reduce OMG, we propose \emph{Joint-Output On-Policy Distillation} (JO-OPD), which distills the model's stronger S2T policy into joint generation using student-generated S2TS trajectories. At each text position, the S2T teacher provides soft targets from a text-only projection of the student's preceding outputs, while the student predicts from the corresponding full interleaved history. A preservation objective further regularizes native non-text predictions. Experiments on Step-Audio-2-mini and Baichuan-Audio-Instruct reveal OMG across two interleaved generation architectures. On Step-Audio-2-mini, JO-OPD reduces OMG from 42.87 to 16.26 percentage points on Spoken-MQA and from 29.72 to 13.04 points on speech-rendered GSM8K, with little change in S2T accuracy and substantially larger reductions than matched SFT baselines. ASR-based evaluation further shows a 7.49-point improvement in spoken-answer accuracy on Spoken-MQA.

Abstract: Neyshekar is presented as an open Persian read-speech corpus designed for coverage of both formal and informal language, named entities, and longer utterances. In version 6, 62,279 validated recordings totalling 99.02 hours are provided from 190 contributors, with 34,541 distinct recorded prompts. The prompt pool was assembled from human-written material, contextualised homographs, and reviewed language-model-generated text. Text entries were normalised with the shekar library, which supports both formal and informal Persian, and every submitted recording was reviewed against a common validation rubric. About 24% of released clips are classified as informal by an automatic classifier; these register labels are not human-validated. Item-level rater labels are provided for reproducible agreement estimation, opaque per-clip contributor identifiers make the speaker-disjoint partitioning auditable and support contributor-clustered uncertainty estimates, and a text-disjoint test subset is included for evaluation beyond previously seen prompts. Per-contributor recording load and reference-free signal quality are characterised for every released clip. Corpus characteristics are compared with Persian Common Voice under shared processing. Utility is assessed through two ASR architectures, three optimisation seeds, WER and CER, and independent evaluation on the public PSRB sample. Against duration-matched Common Voice training at approximately 32 hours, in-domain WER is reduced by 9.5 points for Whisper and 11.6 points for XLS-R, and by approximately eight points for both architectures on the independent PSRB sample. Transfer and mixture benefits are not consistently observed across architectures and training budgets. The corpus is released under CC0; code and data are made available through the project repository at https://github.com/amirivojdan/neyshekar.

Abstract: Deep joint source-channel coding (DeepJSCC) transmits learned semantic features efficiently but can leak sensitive attributes such as gender, race, or speaker identity. We propose LEAPSC (LEACE-in-the-loop privacy for semantic communication), whose core contribution is the integration of in-loop least-squares concept erasure (LEACE) within a variational information bottleneck (VIB) encoder. By periodically refitting the projection operator during training, LEAPSC couples the encoder dynamics to the erasure mechanism, driving attribute-conditional mean differences toward zero within each task-label group on the fitting sample. Additional components, namely conditional value-at-risk (CVaR) tail-sensitive privacy, feature-wise linear modulation (FiLM) signal-to-noise ratio conditioning, and Lagrangian dual ascent, improve robustness across channel conditions and over the high-leakage tail of samples. On CelebA, FairFace, and Google Speech Commands, LEAPSC reaches task accuracy of 0.862, 0.755, and 0.925 respectively, with attacker accuracy at or below the label-only floor on CelebA (0.548 vs. floor 0.580) and within 2 percentage points (pp) of chance elsewhere, improving over an information-bottleneck adversarial baseline (IBAL) at a matched 52-epoch budget by +3.6, +2.5, and +1.3 pp (Welch's t-test, p=0.019 on CelebA).

Abstract: Consonants contribute unequally to whether a word is understood. Given the limited time available for therapy, ranking consonants by contribution to intelligibility helps prioritize intervention targets in motor speech disorders. However, measuring this contribution relies on perceptual studies that are difficult to scale. This paper presents a scalable method that measures consonant contribution using acoustic masking. We silence one consonant at a time in an isolated word and test whether an automatic speech recognition (ASR) model still recognizes the word. We define a consonant's contribution score as the proportion of its masked instances for which the word becomes misrecognized, which we refer to as the mask-induced misrecognition rate (MMR). We validate MMR against two linguistic factors previously reported to correlate with consonant contribution, namely phoneme frequency and functional load. We apply this analysis across four languages, English, Spanish, German, and Czech, using three ASR architectures, MMS (encoder-only), Whisper (encoder-decoder), and Qwen3-ASR (LLM-based). Using partial Spearman correlations, we find that phoneme frequency correlates negatively with MMR while functional load correlates positively. In other words, more frequent consonants are less disruptive when masked, whereas consonants carrying more lexical contrast are more disruptive. Further cross-language analysis shows that consonant rankings are not consistent, indicating that consonant contribution is language-dependent.

Abstract: Neural audio codecs are a fundamental component of modern speech generation systems. While recent codecs achieve increasingly low bitrates, reducing frame rate remains challenging, as each token must preserve more information while maintaining reconstruction quality. We present ZipCodec, a streaming neural speech codec operating at 6.25 Hz and 0.80 kbps with a theoretical latency of 160 ms. Our approach combines large-scale WavLM distillation with a redesigned transformer-based architecture, a scalar spherical quantizer, and a latency-aware streaming decoder. Experiments show that ZipCodec substantially outperforms existing streaming codecs at comparable bitrates in both reconstruction and downstream tasks, while operating at a significantly lower frame rate. Despite its 842M parameters, ZipCodec achieves real-time single-stream inference on a consumer-grade CPU. Demo samples, code and checkpoints are available at https://lucadellalib.github.io/zipcodec-web/.

Abstract: Self-supervised speech foundation models are now used in a wide array of downstream applications, including traditional speech recognition and as the basis for tokens in speech-aware language models. Attempts to understand their usefulness have largely focused on probing their representations' ability to discriminate phonemes and words. However, discriminative ability for words need not imply specialized representation of words per se. Good discrimination of words may be explained by good encoding of word form (phonemes) rather than form-independent word representations encoding identity or syntactic/semantic properties. By partialling out phoneme information using residualization, we show that, in later layers, HuBERT and wav2vec 2.0 do in general learn representations which encode words with reasonable fidelity independently of local phonetic content. We show that this simple approach to disentanglement can enhance higher-order linguistic information in word discovery tasks.

Abstract: Robust speech-to-text translation systems should perform reliably across diverse acoustic conditions, yet practical pipelines lack controllable tools for systematic environment exploration. Large speech models remain sensitive to unseen acoustic conditions, as training data rarely cover the full range of real environments.We present NOPEHYPE, a structured training workflow that combines a controllable environment simulator, coverage-optimal environment reduction on Power Spectral Density (PSD) templates, and a small, interpretable hyperparameter search over simulator knobs. We show that simulator-generated noise achieves performance comparable to balanced realnoise training across Whisper and SeamlessM4T models, provide principled environment prototype sets, and identify practical default simulator configurations from a structured 27-run hyperparameter sweep.

Abstract: Recurrent models provide a natural path to long-context modeling, yet models trained with backpropagation through time (BPTT) often fail beyond their training horizon. Classical analyses emphasize gradients that vanish or explode along temporal paths. However, dense per-token losses can still train a shared recurrent rule despite severe decay, showing that decay alone does not determine whether learning fails. We instead study state credit: the signal through which future losses reach earlier recurrent states before contributing to parameter updates. Accordingly, we intervene directly on state credit and propose Credit Stabilization through Time (CST). During backward propagation, CST locally rescales the state-credit signal to stabilize its norm without rotating the component being corrected, while leaving the forward computation unchanged. Because controlled synthetic tasks and real data exhibit different credit dynamics, we specialize CST to each regime. In both settings, CST improves performance beyond the training horizon, with gains observed at up to 128x the training length.