speech recognition engineers

Abstract: Expanding multilingual ASR systems to new languages is challenging when adaptation must preserve previously supported languages without replaying source-language training data. We study this problem in connector-based Speech LLMs, where a trainable connector maps speech representations into a frozen large language model. We evaluate target-update construction and merging strategies for adapting the connector while preserving source-language performance. In single-target expansion, target-only fine-tuning and standard TIES merging learn new languages but cause severe source-language forgetting, while SSU-Wanda+TIES gives the best source-preserving trade-off. In multi-target expansion, update merging provides stable source retention, whereas recursive expansion can improve target adaptation under favorable orders but is highly order-sensitive. These results show that multilingual ASR expansion requires balancing target acquisition, source retention, and interactions among target-language updates.

Abstract: Transformer-based models have shown strong performance across diverse tasks, but their reliance on self-attention mechanisms often leads to high computational costs. Token sparsification, which reduces the number of tokens processed in transformer layers, offers a promising solution to maintain performance while improving efficiency. However, although widely studied in the vision and language domains, token sparsification remains underexplored in audio, with existing work often limited to a single method, model size, or dataset. This work presents a comprehensive evaluation of token sparsification techniques for audio transformers. We assess a range of token pruning and merging strategies across multiple datasets and model sizes, under both post-training and fine-tuning settings. Furthermore, we introduce a novel sparsification approach that combines attention scores with spectral cues derived from the frequency content of token embeddings. Our experiments show that frequency-aware approaches, particularly those based on spectral flatness and low-frequency energy, consistently outperform attention-only methods. We also find that sparsification based on spectrogram patch intensity is especially effective for longer audio inputs. These results highlight the value of incorporating frequency-domain and input-specific information into sparsification strategies to enhance performance in audio tasks.

Abstract: Spiking Neural Networks (SNNs) promise energy-efficient inference through sparse, event-driven computation, yet remain constrained on continuous, context-dependent tasks such as automatic speech recognition (ASR). A key bottleneck is the Leaky Integrate-and-Fire (LIF) neuron, whose membrane decay erases information across pauses and phonetic boundaries while it must both extract features and retain long-range context. Inspired by cortical separation of nonlinear filtering and somatic integration, we introduce the Spiking Ternary GRU (stGRU): a ternary-weight gated recurrent unit performs temporal filtering, while downstream Resonate-and-Fire (RF) neurons generate the binary spike stream driving the next block's event pathway. This makes binary event transmission a native training constraint rather than a separate deployment-time activation quantization step; together with ternary weights, learned projections avoid dense floating-point matrix MACs. Trained end-to-end on LibriSpeech-960, Small, Medium, and Large stGRUs reach 8.09%, 7.04%, and $6.44 \pm 0.17\%$ WER on test-clean without an external language model. To our knowledge, these are the first matrix-MAC-free spiking networks without a learned ANN sequence decoder to achieve single-digit WER on LibriSpeech. The architecture also achieves 10.88% test CER on Aishell-1 (9.28% on dev) when trained from scratch. Internally, trained networks develop a depth-graded hierarchy: shallow layers act as fast acoustic filters, while deep layers form slower semantic integrators whose latent representations segregate into part-of-speech manifolds and word-level spike events. Deep-layer firing rates follow lognormal distributions with a sparsity gradient, mirroring hierarchical processing observed in human cortex in vivo. Our biologically grounded architecture combines competitive ASR with favorable modeled compute and memory cost.

Abstract: Omni-modal large language models (LLMs) have achieved remarkable progress in audio and video understanding. However, existing models struggle with long-form, multi-speaker videos, especially in understanding the entire video and recognizing who says what and when. To characterize the capability limit of current omni LLMs and guide future model development, we introduce Omnilingual-MsASR, a multilingual, multimodal benchmark for long-form Multi-speaker Automatic Speech Recognition (MsASR) covering six languages, recordings up to 30 minutes, and scenarios with up to 20 speakers. Omnilingual-MsASR provides parallel audio and video inputs to evaluate models' cross-modal consistency and their ability to exploit complementary cues across modalities. Extensive evaluations of leading omni-modal LLMs reveal substantial performance bottlenecks, while their tcpWERs frequently exceed 100%. These findings highlight fine-grained temporal localization, speaker disambiguation, and cross-modal synergy as key challenges for future omni-modal foundation models. Our work is available at https://github.com/QwenLM/Omnilingua-Bench/tree/main/Omnilingua-MSpeaker.

Abstract: Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents. Existing audio backdoors require an attacker to introduce triggers at inference time. We study a different threat model in which the attacker compromises the model during fine-tuning but has no access to the victim's inference-time inputs. We introduce the Speaker-Activated Backdoor (SAB) setting, in which a target speaker's natural voice serves as the trigger. We develop a two-phase fine-tuning method that first makes speaker identity more accessible and then binds the target speaker to an attacker-chosen payload. Compared with a one-phase baseline with the same training budget, the two-phase method substantially improves selectivity and clean performance, and is especially important for content-dependent attacks. SAB achieves 90.2% and 97.0% backdoor success on Whisper-medium and Parakeet, with false-positive rates of 1.7% and 0%, respectively using as little as two minutes of real audio from the target speaker. We further demonstrate content-dependent semantic rewriting and downstream tool-use attacks, where malicious instructions are inserted into 96.9% of target-speaker transcripts. When these transcripts are passed to frontier LLM agents, GPT-4o and Claude (Opus, Sonnet, and Haiku), the injected instruction drives the agent to carry out the harmful action in up to 87.5% of cases. The attack generalizes to held-out recordings, multiple target speakers, and common acoustic degradations. These results establish speaker identity as a practical backdoor trigger that requires no inference-time attacker access and is invisible at the transcript layer.

Abstract: Non-invasive brain-computer interfaces are in practice limited by signal-to-noise ratio, and recent state-of-the-art methods address this limitation by aggregating over repeated trials of the same stimulus. Yet multi-trial aggregation has been treated inconsistently across the literature: some work averages brain signals at the model's input, other work averages predictions at the model's output, and neither treats aggregation as an architectural choice in its own right. We provide the first systematic comparison of aggregation strategies for non-invasive neural decoding, evaluating six paradigms (including logit averaging during training and three attention-based variants we introduce) across three datasets, four tasks, and two group sizes. We find that aggregation choice matters substantially: training-time logit averaging, the simplest method we propose, outperforms all others in 6 of 7 settings. On the phoneme classification setting of the 2025 PNPL competition, it sets a new state of the art, significantly outperforming the competition winner by 17.4 absolute percentage points in balanced accuracy. We further draw a theoretical connection between logit averaging and joint ensemble training, which has been shown to underperform independent ensembling in the deep learning literature; our results call into question recent explanations of why joint ensemble training fails. Together, these findings establish multi-trial aggregation as a key design choice for non-invasive neural decoding and identify training-time logit averaging as a simple, effective default.

Abstract: The recurrent neural network transducer (RNN-T) remains a practical choice for streaming speech recognition, but training requires large intermediate tensors. We present Flash RNN-T, a joint and loss implementation combining specialized kernels with optional tiled recomputation over the full alignment space and vocabulary. Tiled and untiled Flash form the entire measured throughput-memory frontier on medium and long workloads. On H100, untiled Flash achieves 24-39% higher throughput with 26-56% less memory than the fastest baseline on these workloads wherever it fits. Tiling reduces memory further through recomputation. On medium, tiled Flash uses ninefold less memory than that baseline at 16% lower throughput when both use one backward call per batch. On long, it fits where all tested alternatives with that backward schedule run out of memory with 80 GB. These comparisons use each implementation's lowest-memory configuration reaching at least 95% of its peak throughput.

Abstract: In automatic speech recognition (ASR), transcription errors with similar lexical or semantic discrepancies can have markedly different consequences for the situation conveyed by speech. Capturing these differences requires not merely contextual assessment, but an understanding of what the spoken discourse describes, establishes, or seeks to bring about. However, relying on LLM judges introduces computational cost and scoring variability. We introduce SITAR, a situational distortion metric for ASR, realized through a comprehensive end-to-end construction framework. Specifically, this framework aggregates situational severity and cumulative extent into a unified score via a bounded recursive formulation. Using targets derived from this formulation with limited offline LLM support, we train SITAR as a lightweight BERT-family regressor that scores reference-hypothesis pairs in a single forward pass, entirely free of any prohibitive runtime costs. Recognizing that introducing a new metric demands rigorous justification, we formalize a principled validation protocol spanning both extrinsic validity, assessed at macro and micro levels, and intrinsic scorer reliability. The former, evaluated across diverse datasets and models, demonstrates that SITAR reliably tracks broad capability scaling while providing orthogonal diagnostic utility, whereas the latter, examined on a purpose-built controlled benchmark, confirms the operational soundness of the scoring instrument itself. Through systematic validation, we demonstrate that SITAR holds strong potential as a practical, lightweight complement to existing ASR metrics. We will release the full recipes (code, model weights, dataset, etc.) after the internal review to support community research.

Abstract: Speech-capable language models inherit most of their knowledge from text-rich language backbones, and continued model maintenance is likewise dominated by text updates. Knowledge acquired through a text update can nevertheless remain difficult to access when the same query is spoken. We formulate this mismatch as post-update speech access repair, where the knowledge-bearing checkpoint stays frozen and a detachable operator repairs speech-only access in one model pass. We introduce Affine Fields for Interface Repair across Modalities (AFFIRM), which first learns and freezes a base correction for speech hidden states, then adds query-specific corrections with zero mean over the operator-training queries. Paired speech and text states provide the coordinates that condition these corrections on each query. On balanced synthetic-speech PopQA, paired comparisons do not establish a clear advantage for AFFIRM over global SALSA steering or low-rank ReFT interventions under the same frozen single-state constraint. On recorded SLUE-SQA-5 questions with text passages, a full-scale PopQA-fitted AFFIRM operator transfers without refitting and improves ROUGE-L by 0.0941 over its paired No Repair baseline. Low-rank ReFT interventions also transfer positively, whereas the matched SALSA variant falls 0.0707 below its paired baseline. These results show that text acquisition and speech access can diverge after an update and should be verified and repaired separately.

Abstract: Streaming Inverse Text Normalization (ITN) is vital for converting spoken-form outputs from streaming Automatic Speech Recognition into formatted written text. Existing streaming ITN methods rely on hybrid systems combining neural tagging with expert-crafted finite-state transducer rules, limiting scalability across domains and languages. While end-to-end models offer superior scalability by learning directly from data, standard encoder-decoder architectures are inherently non-streaming due to global attention. In this paper, we propose an efficient streaming end-to-end ITN system, adapting a pretrained text-to-text model to leverage its robust linguistic knowledge. To enable streaming, we introduce architectural adaptations, a specialized training strategy, a Read-Tag-Write decoding policy, and inference optimizations. Experiments on Vietnamese datasets show accuracy comparable to non-streaming baselines, outperforming hybrid methods while satisfying real-time latency requirements.

Abstract: Feature distillation is usually expected to add to logit distillation. We find that this expectation does not hold for two common feature losses in compact multilingual automatic speech recognition (ASR), and analyze why through their gradients. The gradient of a projection-based loss is orthogonal to the task gradient, while that of Centered Kernel Alignment (CKA), a similarity-based loss, is weakly aligned with it. CKA lowers the word error rate (WER) on non-English languages, but at the expense of data-rich English. Three loss-level remedies (one trained, two probed) all fall short, and the CKA gradient overlaps with the subspace English uses in the shared encoder by 97.5% to 99.5% (about 81% by chance). Guided by this diagnosis, we propose CAFE (Confined-Adapter Feature distillation), which confines the CKA gradient to adapters that English bypasses, addressing the trade-off in the parameters rather than in the loss. On the Open ASR Leaderboard, CAFE lowers the average WER from 8.55% to 8.02%, beats an adapter-only capacity control at 8.51%, and makes English more stable across checkpoints. At 254M parameters, it trails the 0.6B teacher by 2 WER points and is competitive with much larger models across 48 test sets in 24 languages. We use 564k hours of training data and state-of-the-art models to keep the findings relevant to practice.

Abstract: Audio memory must recognise recurring sources across acoustic changes while keeping their attributes up to date. Overall accuracy does not reveal how these demands interact. We introduce Audio Memory Operations (AMemOps), a benchmark for latest-attribute retrieval, preservation of unchanged attributes and absent-source rejection in long audio streams. Paired controls vary recordings, training identities and earlier events to examine their effects on updating. Training on other recordings of the test instruments increases a reference delta-rule memory's update advantage over matched attention by 13.4 percentage points. At retrieval delays of 120\,s, matched histories separate benefits of consistent repetition from costs of conflicting earlier values, while additive-memory controls identify benefits of erasure.

Abstract: Supervised fine tuning (SFT) is commonly monitored through training and evaluation loss, but loss does not fully characterize how model functionality evolves along the same trajectory. We study this gap through dialect adaptation of Qwen3-ASR, MiMo-Audio, and FireRed-LLM across learning rates from $10^{-5}$ to $10^{-3}$. Character error rate (CER) reveals pronounced transient collapse and recovery whose severity and temporal structure are only weakly reflected by loss. Bidirectional layer patching reveals strong cross checkpoint functional incompatibility during recoverable collapse, with model dependent depth profiles that reorganize during recovery. Pairwise gradient similarity alone does not reliably indicate optimization stability, whereas raw gradient spectral analysis provides a clearer structural distinction between stable and unstable optimization: successfully converged runs approach a broad and temporally stable spectral energy distribution, while failed runs remain spectrally concentrated or exhibit persistent fluctuations. Finally, AdamW update transfer bypasses pronounced early collapse, yet substantial Mandarin degradation still emerges later in training. These results motivate three separable SFT dynamics: collapse, repair, and drift, and show that transient collapse is not necessary for persistent source domain degradation.

Abstract: Omnimodal large language models (OLLMs) are susceptible to cross-modal hallucinations under audio-visual interference. We find that, for a substantial fraction of incorrect joint predictions, the question-relevant unimodal branch can still recover the correct answer, suggesting that task-relevant evidence can remain recoverable from the relevant modality even when it is not reflected in the joint prediction. We refer to this recoverable failure pattern as \emph{modality misrouting}. We propose CaRe (Conflict-aware Residual Intervention), a training-free inference-time method for identifying and correcting modality misrouting. CaRe uses a text-only modality probe to identify a candidate modality-specific counterfactual branch, and flags potential misrouting when its prediction conflicts with the original joint prediction. For such cases, CaRe computes a residual direction from the discrepancy between the counterfactual and fused representations and uses it to steer the original multimodal state toward the relevant modality evidence, while retaining both modalities in the original audio-visual inference path. The intervention strength is adapted based on the confidence of the counterfactual branch. Experiments on CMM and AVHBench across two OLLMs show that CaRe consistently outperforms the evaluated inference-time hallucination-mitigation baselines.

Abstract: Continuous latent representations underlie recent diffusion-based lyrics-to-song and text-to-music systems, yet aligning lyric conditions with these latents remains difficult: existing systems resort to generator-side machinery such as sentence-level timestamps, auxiliary alignment objectives, or semi-autoregressive decoding, while the latents they predict are still produced by autoencoders trained purely for reconstruction. We address this difficulty from the representation side. We propose MSVAE (Music Semantic Variational Autoencoder), which structures its continuous latent around two complementary semantic streams: work-level compositional content and lyric-level linguistic-vocal content. Partitioned dual-stream alignment makes both streams explicit by aligning disjoint latent subspaces with frozen lead-sheet and ASR references, while a semantic-first gated residual posterior derives its base from a frozen music-SSL encoder and admits acoustic information only through gated, bounded corrections. MSVAE achieves competitive stereo reconstruction and stronger MARBLE probing than continuous-autoencoder baselines, and improves downstream lyrics-to-song and text-to-music generation. Notably, with only an untimed phoneme prefix and no lyric timing, MSVAE yields the lowest phoneme error rate, whereas all baseline representations require sentence-level alignment to produce intelligible vocals.

Abstract: Audio source separation models must preserve fine spectral structure while keeping the cost of the separator tractable. Band-split architectures reduce this cost by manually partitioning the spectrum into predefined frequency bands and encoding each band as a token, but the same boundaries are used for every mixture. Because informative spectral structure varies with source content, fixed boundaries can merge discriminative components or spend tokens where little resolution is needed. We introduce Dynamic Band Splitting (DBS), an end-to-end framework that adapts boundary locations and band widths to each mixture under a fixed band-token budget. DBS predicts content-conditioned boundary scores within frequency blocks, selects a prescribed number of boundaries using hard Top-$K$, and applies content-dependent pooling to obtain compact band tokens. After separation, it smooths and expands these tokens to the original STFT grid, refines them along frequency, and decodes complex masks with a frequency-shared head. DBS improved separation over the corresponding static models on DnR v2, MUSDB18-HQ, and EchoSet, with DBS-BSRNN achieving relative gain of 29.0% in average SI-SDR on DnR v2. These results and boundary analyses on DnR v2 show that content-adaptive frequency partitioning benefits multiple separator backbones under a fixed band-token budget. Code and audio demos are available at https://cslikai.cn/Dynamic-Band-Splitting

Abstract: Target speaker extraction (TSE) is a key frontend for enabling audio large language models (LLMs) to recognize and understand a designated speaker in overlapping speech. However, multilingual mixtures exhibit heterogeneous extraction demands, causing fixed-depth systems to either under-process challenging inputs or waste computation on inputs that have already converged. Existing TSE frontends generally apply uniform computation, while Audio-LLM adaptation methods neither explicitly model language-conditioned extraction demands nor preserve plug-and-play compatibility with frozen downstream models. We therefore propose LATTE, a lightweight 5.2M-parameter frontend that combines an acoustic Difficulty Estimator, a task-adaptive phonology-informed prior learned from multi-depth TSE gains, and an Adaptive Hourglass Extractor to route each mixture without requiring language identity at inference. On PolyExtract, LATTE achieves a macro SI-SNRi of 15.07 dB at 0.30 GFLOPs, retaining performance within 0.09 dB of a 0.53-GFLOP fixed-depth model while reducing computation by 43.4%, and lowers the average WER from 16.8% to 14.6% relative to TF-GridNet. By jointly modeling sample-level extraction demands and phonological structure, LATTE provides a practical framework for quality-aware, compute-efficient multilingual TSE and offers a scalable path toward robust Audio-LLM deployment.

Abstract: Earnings call transcription presents a realistic continual learning scenario where Automatic Speech Recognition (ASR) systems are updated with new recordings but risk catastrophic forgetting of previously learned financial speech patterns. Adapting Whisper-small from a general-domain speech corpus to the Earnings22 benchmark reveals the opposite: all evaluated methods exhibit improvement on held-out financial speech from the same domain during adaptation rather than forgetting. This task complementarity makes regularization counterproductive. Nevertheless, when regularization is warranted, a layer-wise mini-Kronecker-Factored Approximate Curvature (mini-KFAC) approximation seems to achieve lower word error rates (WERs) and more stable optimization than standard diagonal Elastic Weight Consolidation (EWC). The results suggest that task complementarity should be assessed before applying continual learning constraints, and prefer mini-KFAC EWC over diagonal EWC when needed.

Abstract: Softmax-based multimodal emotion recognition methods often assign high confidence even when sensor data are corrupted or contradictory, for example when facial expressions suggest positive valence while vocal prosody indicates negative arousal. We present an Evidential Multimodal Transformer with Dempster-Shafer Fusion (EMT-DS) that addresses this problem through formal uncertainty quantification. EMT-DS models predictions as Subjective Opinions (Dirichlet distributions), separating epistemic uncertainty (lack of evidence) from aleatoric uncertainty (sensor noise). Modality-specific transformer encoders produce belief mass distributions that are fused with Dempster's Rule of Combination, which normalizes the joint belief mass by a conflict coefficient 1/(1-c), so that contradictory evidence is explicitly quantified as conflict and the fused prediction becomes diffuse rather than being forced into an unjustified consensus. Under subject-independent and speaker-independent evaluation on four benchmarks, EMT-DS achieves 84.3% accuracy on IEMOCAP versus 81.7% for the best baseline (p < 0.01), 82.4% on DEAP under leave-one-subject-out evaluation versus 78.1%, and 96.8% on the highly imbalanced KEMDy20, and it improves accuracy on CMU-MOSEI by 2.4 points over the strongest evidential baseline. In controlled ablation, it reduces Expected Calibration Error from 0.153 (softmax transformer) to 0.042. It maintains 71.2% accuracy under severe Gaussian noise on audio features (sigma = 1.0), where attention baselines drop to 45.3%, and degrades gracefully when entire modalities are absent. The fusion step has O(MK^2) complexity, compared with O(M^2 d^2) for cross-modal attention. The learned conflict metric also flags likely errors: reviewing 20% of samples captures 80% of misclassifications, which supports selective prediction in clinical and safety-critical systems.

Abstract: German is a morphologically rich language whose syllable structure is exceptionally well-predicted by the Knuth--Liang hyphenation algorithm. We ask whether phonologically informed tokenization can serve as a competitive target for end-to-end speech recognition. We compare three tokenizer families on the Omnilingual ASR wav2vec 2.0 backbone fine-tuned with CTC: the pretrained multilingual character inventory, a data-driven Byte-Pair Encoding (BPE) over orthography, and phonologically informed units from Pyphen syllabification and grapheme-to-phoneme conversion. Across 40 fine-tunes, we evaluate on three German test sets spanning orthogonal shifts: in-domain read speech, dialectal spontaneous speech, and standard-German spontaneous speech. In-domain, all phonologically informed tokenizers match BPE and the multilingual character baseline on both WER and CER. Under domain shift the picture splits along vocabulary size rather than the linguistic axis of variation: at small vocabularies, syllable-aware tokenization improves on dialectal speech, where phonetic surface forms vary but syllable structure is preserved, and stays ahead on spontaneous speech, where new word-forms violate vocabulary closure. A phoneme-level confusion analysis further shows that all tokenizers commit the same canonical function-word errors, indicating that the acoustic encoder, not the tokenizer, dominates the error topology. Our findings suggest that tokenizer choice may depend on the vocabulary budget as much as on the distribution shift expected at deployment rather than reducing to a single universal optimum.

Abstract: End-to-end speech parsing, a task recently proposed, consists in predicting both the transcription and the syntactic tree for a spoken utterance. Existing architectures for speech parsing often utilise intermediate neural networks. In this work, we examine the effectiveness of intermediate neural networks (NN) for parsing, and, specifically, what role do they play. We introduce a simpler end-to-end architecture for speech parsing, where we remove these intermediate NN units, reducing the parameters by 12%, while achieving comparable or better performance than prior method on both automatic speech recognition (ASR) and parsing. We demonstrate that intermediate NN units help reduce the representational gap when the pre-trained encoder is frozen. We do a comprehensive evaluation of speech parsing on French, and medium-low resource languages Slovenian and Naija. We further investigate the impact of the training data size and intermediate layers of the pretrained speech encoder on speech parsing.

Abstract: We present InterView-C, a German multimodal corpus of 27 survey interviews conducted entirely in virtual reality, with both interlocutors represented by avatars. The corpus aligns spoken interaction with synchronized behavioral data, including gaze, head and body movement, facial behavior, hand and finger tracking. Its reference transcripts and linguistic annotations provide a reliable interface between this multimodal spoken interaction and predominantly text-based NLP methods. This interface is important because automatically transcribing speech can distort linguistically relevant information, while downstream models trained on existing resources may additionally face transfer challenges when applied to transcribed spoken data. InterView-C therefore provides word-timed and manually post-edited verbatim transcripts for all 54 recordings, interview-item timings, questionnaire responses and negation cue and scope annotations for 1,422 sentences, 1,398 of them doubly annotated (α=0.87 for cues; α=0.81 for scopes). We demonstrate both challenges empirically: nine open-weight ASR systems disproportionately misrecognize short closed answers and number words, while negation models trained on existing corpora show lower and highly variable performance on our transcribed interviews than a model trained on the InterView-C annotations. InterView-C thus enables linguistic analyses of spoken interaction while retaining their alignment with rich multimodal behavior.

Abstract: Delayed Streams Modeling (DSM) enables LLM-based streaming automatic speech recognition (ASR) by aligning acoustic and text streams on a common timeline. DSM adds the padding token <p> and the word-start token <w> to the LLM vocabulary and predicts them together with normal text tokens using the same softmax. We first show that <w> can be removed while maintaining competitive recognition performance. Based on this result, we propose Factorized DSM (F-DSM), which separates the waiting probability for <p> from the distribution over the original LLM vocabulary. This factorization removes ASR-specific tokens from the text prediction space and allows the large-vocabulary softmax to be skipped on waiting steps. Experiments on the Corpus of Spontaneous Japanese and LibriSpeech show that F-DSM achieves better recognition performance than DSM. It also greatly reduces GPU memory use while maintaining similar training throughput, provides a small inference speed improvement through softmax skipping, and reduces the degradation in text-only perplexity observed with DSM.

Abstract: Aggregate accuracy cannot reveal which utterances are locally vulnerable or how internal activity changes when labels remain stable. We retain every prediction for 725,070 adjacent-bin, one-count changes around 100 validation utterances of a frozen SpikeSCR-based classifier. The canonical native-horizon GPU singleton path reaches 86.0836% validation accuracy. Equal-source expected accuracy under a uniformly chosen neighbor rises from 84.00% to 84.54%, although 13 of 84 initially correct sources admit adverse neighbors. Five sources carry 93.21% of adverse moves. Margin-guided and surrogate-gradient searches miss sparse cases at fixed query budgets; a post hoc gradient prefix finds all 13 at 73.45% of the census of initially correct sources. Source-matched traces and clean-state interventions distinguish internal change from harmful direction and recoverability. Two count-readout replicas have similar validation accuracies but a 5.21-fold class-change gap concentrated in four sources. Controlled query/key source isolation eliminates 531 batch-order label changes and restores bitwise order invariance across all 9,981 validation score vectors. Isolated batch predictions reproduce padding-matched singleton labels on 9,980 of 9,981 inputs. Paired CPU/GPU replay of all 25,820 class-changing neighbors gives 99.8993% label agreement and preserves all 13 vulnerable sources and their fixed witnesses. Complete source-conditioned maps distinguish vulnerability incidence, concentration, internal change, and execution dependence that aggregate accuracy leaves unresolved.

Abstract: Self-supervised learning (SSL) speech models are accurate but large. Knowledge distillation compresses them, but the student loses the teacher's noise robustness. Correlation-based distillation addresses this with two terms: cross-correlation aligning student and teacher's representations, and self-correlation decorrelating the student's features. Both the original method and De'HuBERT credited the self-correlation term without isolating it. We show the opposite. On LibriSpeech-100 distillation with held-out CHiME-3 noise at 10\,dB, an unbiased full-dimensional probe, a Pearson-variance decomposition, a same-noise causal control, and a per-dimension analysis identify the cross-correlation diagonal as the mechanism that encourages noise invariance, lowering noise-classification accuracy from 76.98\% to 55.02\%, whereas adding the self-correlation term back leaves it at 55.55\%. The self-correlation term instead reorganises the feature space and improves accuracy on the clean downstream tasks, but removes essentially no noise. Across nine speech and music tasks this yields a concrete design rule for distillation: weight the cross-correlation term for noise robustness, tune the self-correlation term for better downstream performance, and sample teacher and student noise independently.

Abstract: Improving joint audio-visual reasoning in Omni Large Language Models typically incurs substantial data construction and training costs. Our diagnostics reveal multi-hop reasoning difficulties despite correct answers to all corresponding single-hop questions and suggest partial decoupling in the local optimization of perception and reasoning objectives. This motivates post-training with different emphases on these capabilities. Text-only reasoning training yields gains across data sources, model scales, and families. With the best-performing text-only configuration, supervised fine-tuning followed by reinforcement learning (RL) raises Qwen2.5-Omni-7B's geometric mean of nine reasoning scores by 25.83% over the base model, outperforming the complete native audio-visual route with 56.6% fewer GPU-hours. Training on data synthesized entirely by a text-only LLM raises this geometric mean by 21.01% without audio-visual data in construction or training. However, text-only training degrades perception. We therefore propose a text-centric post-training paradigm: text-only training provides the main reasoning optimization, and reduced-data native audio-visual RL then refines perception. Refinement uses about 90% fewer input tokens than full-data audio-visual RL, restores perception above the base level, and retains 93.5% of the best-performing text-only pipeline's reasoning gain.

Abstract: Pitch tracking systems typically couple a per-frame fundamental frequency ($F_0$) estimator with a temporal smoothing stage to obtain continuous trajectories. Conventional Viterbi smoothers enforce first-order continuity but lack long-term temporal awareness and could lock into octave errors across corrupted frames. We propose Relative Interval Networks (RIN), a trajectory smoothing framework that reconciles per-frame pitch estimates with data-driven multi-hop pitch differences. We extract robust relative pitch intervals across arbitrary frame offsets using Variable-Q Transform cross-correlation. We formulate pitch smoothing as an $L_1$-norm optimization problem and prove its equivalence to a minimum cost circulation problem, solved efficiently via linear programming. Evaluations across speech, singing, and instrumental datasets show that RIN substantially improves weak estimators, matches or outperforms Viterbi decoding at a comparable computational cost, and provides superior robustness under certain acoustic degradation.

Abstract: Demographic fairness gaps in automatic speech recognition are almost always reported from a single training run. We fine-tune the Q-former projector and LoRA adapters of a speech LLM at five audio compression factors and six random seeds, holding the encoder, base decoder, data and decoding fixed, and evaluate every run on Common Voice and Fair-Speech. At 460 h of clean LibriSpeech, the seed moves fairness metrics more than compression does on most demographic axes. A balanced 3x3 decomposition attributes 85.3% of the variation in Fair-Speech ethnicity normalized gap to the seed against 8.3% to compression (p = 0.009), though compression explains more on age and gender. Held-out LibriSpeech word error rate spreads by 0.04 points across those seeds while Common Voice spreads by 8.57, so these are not failed runs, and the effect survives controlling for accuracy and dropout. Scaling and diversifying the adaptation set to 960 h damps the effect but does not remove it. On Fair-Speech ethnicity, two single-run systems must differ by more than 0.30 in normalized gap to exceed seed variability.

Abstract: A large audio language model (LALM) turns a minute of speech into 750-1,500 tokens and prefills every one. Image-token pruning often cuts after the language model's first layers, where image tokens draw little attention. Audio tokens draw much more attention there, and their ranking is still far from final, so audio needs a ranking before the language model runs. Surprisingly, the attention an audio token will receive across the language model is already linearly predictable from its encoder output, before the language model runs. A linear map, fitted in closed form without labels, predicts this all-layer attention ranking at $ρ\geq .69$ on eleven of thirteen LALMs. Our method, Triage, cuts audio tokens by this prediction and, on multiple choice, cuts again at layer 2, correcting the prediction with the attention observed there. Triage sets its compression without labels, under two budgets that limit how far its output may differ from the model's own full-audio output. At the conservative budget, its word error rate and accuracy stay within .04 of full audio. At the aggressive budget, Triage beats every baseline in all twelve transcription cases. On multiple choice, at 2.2-5x compression, it outperforms DART, the strongest baseline on average, by .043 in mean accuracy. Because it cuts before the language model, it raises the audio that fits in Qwen2.5-Omni-3B's context window from 21.8 to about 62 minutes. At its most compressive point, Triage lets one GPU serve 4x as many concurrent 5-minute streams of that model. Project page: https://audio-triage.github.io

Abstract: Speech-LLMs are expensive to run, making compression important for real-world deployment. However, compressed models are usually selected using aggregate word error rate (WER), which can hide how pruning affects different demographic groups. In this work, we systematically study the effect of audio encoder pruning on SLAM-ASR for different demographic groups. Using the Fair-Speech and Common Voice datasets, we found that the pruning does not affect all demographic groups equally; the gap between best- and worst-performing groups increases in fold. These disparities appear across all three encoder scales, but only the largest model initially hides them behind aggregate WER. LoRA adaptation improves WER for every group, but benefits groups already performing well more strongly and widens for certain groups. On Common Voice English, Danish, and Dutch, accent gaps persist but do not clearly widen, showing that the fairness effects of pruning vary across datasets and must be measured directly. Our findings suggest that for pruned models, deployment decisions should include per-group WER, with the worst-performing group's error rate as an explicit criterion.

Abstract: Reasoning about temporal structure of audio recordings requires Large Audio Language Models (LALMs) to associate sound events with their temporal position. Understanding the underlying mechanisms is a first step toward diagnosing failures and identifying model components that may need improvement. Using mechanistic interpretability, we investigate how temporal information is represented and bound to sound events in three open-source LALMs. We find that across all three, event-specific location becomes concentrated in event name representations at intermediate modality integration layers. These representations encode coarse event position along a low-dimensional, curved relative time trajectory. Steering event name representations along this trajectory systematically shifts before/after beliefs, providing evidence that these representations contribute to coarse temporal reasoning. In contrast, the same interventions do not reliably shift predicted onset timestamps, suggesting that coarse temporal reasoning and precise metric event localization rely on distinct mechanisms.

Abstract: A central aim of neuromorphic computing is to provide a viable alternative to highly energy-intensive Transformer-based AI. However, efficient alternatives struggle to capture the set of qualities that have secured the Transformer's status as the de facto standard in sequence modelling. Any realistic contender must be data-adaptive, able to capture long-range dependencies, and GPU-parallelisable, but also non-linearly recurrent to enable complex reasoning. Based on evidence suggesting the auditory cortex operates on fixed timescales, this work proposes the ADaptive with Prescriptive Timescales Network (ADPTNet) as a potential solution to achieving all four properties simultaneously. ADPTNet is built around local topological conjugates, obtained by a novel combination of linear attention and Riemannian optimisation, applied to static global dynamics. This enables non-linear yet predictable long-term behaviour. Dynamical systems theory proofs provide theoretical guarantees for the parametric control of ADPTNet's timescales (its Lyapunov spectrum). ADPTNet improves performance on Selective Copying over Hawk, the existing method balancing long-range memory and adaptability, while also improving state tracking over linear SSMs like Mamba. On sequential CIFAR-10, ADPTNet matches linear SSM accuracy and outperforms existing selective models (incl. the Transformer), using fewer parameters. We also introduce a neuromorphic SpikingADPTNet, which achieves a new state-of-the-art accuracy on the Spiking Speech Commands dataset ($83.56\%\pm0.15$). Finally, ADPTNet's constant timescales enable two efficient, Jacobian-free extensions to the DEER parallel simulation algorithm (Conv and Forward DEER) that retain the same average convergence. Conv DEER adds no computational overhead beyond the network's forward pass and enables non-linear RNN parallelisation via iterated convolutions for the first time.

Abstract: Decoding imagined speech is an important goal for brain-computer interfaces but remains challenging due to weak neural responses, low signal-to-noise ratio, and limited imagined-speech datasets. Language models provide strong contextual cues for text prediction, but how much they can help neural decoding and whether their contribution differs for decoding perceived and imagined speech remains unclear. To investigate this, we use a paired listened-imagined MEG dataset and incorporate language-model information at two stages. First, we train a contrastive neural decoder that aligns MEG representations with acoustic and contextual language representations, improving cross-subject word decoding for both listened and imagined speech. Second, at inference, we introduce a neural-constrained beam-search framework that combines neural evidence with language-model next-word probabilities. We find that imagined-speech decoding benefits more from the language model than listened-speech decoding. For Imagined speech, the best-performing balance between neural and language-model evidence shifts toward the language model, and the gain over neural-only decoding is larger. Together, these results suggest that language priors are most useful when neural evidence is weaker, making them particularly valuable for imagined-speech BCIs.

Abstract: Open-vocabulary keyword spotting (KWS) must detect arbitrary keywords without retraining. Cross-attention-based models achieve state-of-the-art performance but require pairwise interaction between the audio and each keyword, recomputing the fused representation for every audio--keyword pair. Embedding-based models avoid this computational cost through independent encoding and similarity scoring, yet remain inferior on challenging benchmarks such as LibriPhrase-hard. We show that this gap is primarily a training issue rather than a limitation of model capacity. While contrastive learning encourages correct keywords to rank above negatives, it does not explicitly calibrate absolute similarity scores, which are critical for applying a fixed threshold across keyword detection. We propose Calibrated, Order-aware Detection KWS (CORD-KWS), a training framework that introduces two complementary objectives while retaining the same encoders, cosine scoring, and O(1) keyword enrollment: a calibrated detection head that supervises absolute scores, and a frame-level CTC objective that preserves phonetic order information. CORD-KWS achieves EERs of 0.43% and 9.64% on LibriPhrase-easy and LibriPhrase-hard, respectively, outperforming existing embedding- and cross-attention-based models.

Abstract: We introduce target-speaker unlearning ASR (TSU-ASR) task in a fully end-to-end framework for multi-speaker ASR and diarization. Given a multi-speaker utterance and a set of opt-out speakers who do not wish to have their speech transcribed, the task requires an ASR system to transcribe all speakers except the opt-out ones, while still indicating when those speakers are active. As a first step towards tackling this task, we introduce a novel, light-weight Enrollment-Conditioned Gating (ECG) module attachable to a frozen dual-stream speech LLM that enables ASR for new opt-out speakers dynamically during inference, even those who were not seen during initial ECG training phase. Our experiments on both AMI (English) and AliMeeting (Mandarin) datasets show that speech transcription accuracy for corresponding opt-out words or characters falls from 72.3% to 48.2% and from 73.6% to 27.3%, respectively, while retained speakers' transcription error rates maintain more or less the same. Our approach provides a practical solution for modern video conferencing platforms, allowing speakers to dynamically opt-out from automated AI transcriptions without forcefully leaving the meeting sessions, enabling a privacy-preserving interface for potentially millions of online meetings daily.

Abstract: Recent advances in speech language models have improved automatic speech recognition (ASR) for long-form audio. However, accurately and consistently transcribing domain-specific terminology remains challenging. Motivated by the world knowledge and contextual capability of large language models (LLMs), we propose Agentic-GER, an LLM-based agent for terminology correction in long-form speech. The agent uses global context from the full transcript to identify suspicious terms and resolve ambiguous hypotheses. It selectively re-transcribes the source speech to check candidate corrections, and uses accepted edits to guide subsequent decisions. Experiments with four LLMs and two ASR systems on GigaSpeechBench show consistent terminology improvements in both Chinese and English, with and without thinking. On Chinese speech, Agentic-GER achieves up to a 36.8% relative reduction in biased character error rate (B-CER) over the Whisper baseline.

Abstract: Pruning large pre-trained transformer-based ASR models such as OpenAI's Whisper has seen great adoption, as pruning the decoder led to significant end-to-end transcription speedups. For instance, the {\tt whisper-large-v3-turbo} variant reduced the decoder from 32 to 4 layers, while Distill-Whisper similarly reduced the decoder to only 2 layers. Although some attention has been put towards reducing the size of the encoder, no approach has seen wide adoption. This could be due to the need for custom inference implementations to take advantage of the compressed model. We present an approach that ranks encoder layers by the leave-one-layer-out change in Word Error Rate (WER). The six layers that cause the least change are removed, corresponding to $18.5\%$ of the encoder stack. The pruned model requires no custom inference code as it is simply a more shallow encoder with fewer layers. We further distill using unlabeled monolingual speech data to recover performance degradation caused by the zero-shot layer pruning. Mean WER across four languages increases to $20.1\%$ after distillation, compared to $21.9\%$ zero-shot, going from a baseline of $18.2\%$. We release all of our code (https://github.com/rasgaard/whisper-encoder-layer-prune) and the pruned model (https://huggingface.co/rasgaard/whisper-large-v3-turbo-encoder-pruned).

Abstract: Short-to-long training is a simple curriculum for speech models, but its gains can be difficult to interpret. In speech token language models, length-based training can change the shuffle policy, batch composition, token retention, and token weights under batch-mean loss. We disentangle these factors through matched comparisons. In the tested settings, short-to-long ordering shows no independent benefit when batch composition and token exposure are fixed. First-epoch grouping lowers perplexity for Mimi under batch-mean loss, but this gain is not observed under token-balanced loss. The cross-tokenizer results are consistent with a link between chunk-length variation and token weighting. This work provides a systematic analysis protocol for studying length-based training in variable-length speech models.

Abstract: Speech-to-text engines are extremely needed nowadays for different applications, representing an essential enabler in human-robot interaction. Still, some languages suffer from the lack of labeled speech data, especially in the Arabic dialects or any low-resource languages. The need for a self-supervised training process and self-training using noisy training is proven to be one of the up-and-coming feasible solutions. This article proposes an end-to-end, transformers-based model with a framework for low-resource languages. In addition, the framework incorporates customized audio-to-text processing algorithms to achieve a highly efficient Jordanian Arabic dialect speech-to-text system. The proposed framework enables ingesting data from many sources, making the ground truth from external sources possible by speeding up the manual annotation process. The framework allows the training process using noisy student training and self-supervised learning to utilize the unlabeled data in both pre- and post-training stages and incorporate multiple types of data augmentation. The proposed self-training approach outperforms the fine-tuned Wav2Vec model by 5% in terms of word error rate reduction. The outcome of this work provides the research community with a Jordanian-spoken data set along with an end-to-end approach to deal with low-resource languages. This is done by utilizing the power of the pretraining, post-training, and injecting noisy labeled and augmented data with minimal human intervention. It enables the development of new applications in the field of Arabic language speech-to-text area like the question-answering systems and intelligent control systems, and it will add human-like perception and hearing sensors to intelligent robots.

Abstract: Attention encoder-decoder (AED) Speech Foundation Models achieve strong ASR performance but can generate acoustically unsupported text when inputs contain no speech, weak acoustic evidence, or unreliable transcription. We propose AURA: Activation-editing with Uncertainty-Routed Adaptation, an ultra-efficient representation-editing method that freezes the pretrained model and applies sparse scale-and-shift edits to decoder cross-attention heads. AURA dynamically routes edits using cross-attention uncertainty features that capture over-concentration, diffuse attention, and abrupt frame shifts. We evaluate AURA on four datasets spanning non-speech hallucination and speech grounding stressors, including imperfect-label child speech, imperfect-label adult speech, and disfluent speech. On non-speech audio, AURA reduces hallucination rate from 89.18% to 1.94% without prior hallucination-head identification. On imperfect-label corpora, AURA approaches LoRA WER while using roughly 500x fewer trainable parameters. Sensitivity analysis and qualitative cross-attention examples are consistent with AURA's uncertainty-routed editing behavior, supporting dynamic activation editing as a practical path for grounding AED speech models.

Abstract: We present the Spoken Wikipedia Presentation Corpus, an extension of the Spoken Wikipedia Corpora featuring LLM-generated slide decks for multimodal ASR. Slides are created from LLM-segmented sections using a hybrid pipeline that combines LLM-based content planning with rule-based design decisions. For each section, an LLM generates a slide title, bullet points, a takeaway message, and a visual description that is used to create an illustration. Rule-based matching then selects layouts, themes, and styles to produce the final slides. A vision LLM extracts slide text as Markdown. We evaluate multiple ASR and spoken language models (SLMs). The best model achieves an average micro-WER of 10.23% and an average micro-CER of 6.48% on audio-only inputs. English yields the lowest error rates, followed by German and Dutch, while performance declines across lower-resource languages. Although audio-only baselines are strong, multimodal zero-shot prompting of omni models remains challenging. The aligned slide, text, and audio data show a strong potential to improve recognition through cross-modal context.

Abstract: Real-time multichannel speech extraction (MSE) under diffuse noise conditions is an important task with a wide range of applications, such as speech recognition and hearing aids. In this paper, we propose online algorithms for independent low-rank matrix analysis (ILRMA) and rank-constrained spatial covariance matrix estimation (RCSCME). Previously, we proposed a real-time extension of the RCSCME-based method: an MSE method based on ILRMA and RCSCME using the blockwise batch algorithm. However, it assumes that the spatial characteristics are stationary within a single batch, and thus, in dynamic situations where the target speaker moves, its performance may degrade. To address this problem, we derive the online algorithms for ILRMA and RCSCME in the following three steps. First, we formulate framewise cost functions for ILRMA and RCSCME on the basis of maximum weighted likelihood estimation. Second, we derive the update rules for the framewise cost functions on the basis of auxiliary-function techniques. These naive update rules are computationally costly for real-time execution on a practical machine. Thus, we finally derive the online algorithms by approximating some intermediate parameters with their estimates. Furthermore, we propose stabilization and further acceleration techniques for these online algorithms. In experiments, we simulate situations where a target speaker is stationary or moves and show that the proposed method achieves superior speech extraction performance compared with conventional methods. In addition, using real-world recorded signals, we demonstrate the effectiveness of the proposed method in practical scenarios.

Abstract: Question answering has advanced rapidly with large language models, but predominantly for high-resource languages, in both text and spoken settings. Spoken question answering (SQA) benchmark for Telugu remains unexplored, and the reliability of automatic evaluation in this setting remains unquantified. We introduce VākQA, a Telugu SQA benchmark of 2,001 factoid question-answer pairs across six domains, with 2.53 hours of speech audio, bilingual transcriptions, and human-verified reference answers. We first validate evaluation methods against human judgements: Gemini-as-a-judge best approximates human ratings but is non-uniformly strict, while open-weight judges systematically penalize correct Telugu answers that differ in surface form from the reference. Using this validated setup, we benchmark proprietary and open-weight models across input modality, language, and domain. We observe that Telugu phrasing retains cultural specificity that is lost in translation, speech input introduces phonetic confusions that alter question meaning, and cascaded ASR-MT errors compound progressively. VākQA is publicly released.

Abstract: Contextual ASR can benefit from semantic cues or from target words explicitly provided in the context. We introduce HearInContext, a Mandarin--English benchmark that pairs shared synthetic speech with assistant replies supporting different interpretations. The benchmark comprises 3,764 semantic test cases built around homophones. Implicit contexts exclude candidate words; explicit contexts name the target. No-context and unrelated-context controls measure the benefit of relevant history and sensitivity to irrelevant history. Context-capable models benefit from implicit cues but achieve higher target recall with explicit hints. Fine-tuning Qwen3-ASR-1.7B improves implicit-context target recall by 11.0 and 11.5 percentage points in Mandarin and English, respectively, while absolute CER/WER changes on AISHELL-1 and LibriSpeech remain below 0.1 percentage points. Gains extend to explicit conditions excluded from fine-tuning and to Mandarin hotword recognition on real recordings.

Abstract: While Large Language Models excel in natural language processing, efficiently extending their capabilities to spoken input remains a significant challenge. Existing methods for building SpeechLLMs often rely on computationally expensive full-model fine-tuning, or employ parameter-efficient projectors that suffer from inefficient token sequence lengths and costly full-model supervision. In this paper, we introduce Aligned Continuous Integrate-and-Fire, a highly efficient framework for zero-shot speech processing. Our method dynamically compresses continuous acoustic frames into the exact discrete token length of the target text utilizing explicit Dynamic Time Warping alignments. This allows our initial training stage to establish a robust acoustic-to-semantic bridge using lightweight distance metrics, entirely bypassing the computationally expensive LLM forward pass. For subsequent fine-tuning, we propose a memory-efficient knowledge distillation objective that targets a single LLM layer, performing competitively with full-model cross-entropy training at a fraction of the computational cost. Through extensive evaluations on Automatic Speech Recognition and Speech Translation, we demonstrate that our method achieves superior performance compared to prior parameter-efficient baselines.

Abstract: This paper establishes a perfect-reconstruction (PR) interpretation of AddUNet and its full-rate realization, and introduces a Residual Full-Rate PR architecture for task-directed representation learning. The survivor--skip structure of a constrained additive U-Net is shown to be exactly equivalent to a critically sampled multirate PR filter bank. The full-rate formulation removes the complementary-subband restrictions of the critically sampled system while preserving PR. A Residual Full-Rate PR architecture is then proposed to progressively route task-irrelevant, nuisance, or redundant structure away from the task-facing survivor while retaining the routed information explicitly. Exact reconstruction is guaranteed for arbitrary shape-compatible linear or nonlinear routing operators, without requiring invertibility, a matched synthesis bank, reconstruction loss, or learned decoder. The resulting architecture decouples representation design from reconstruction design: conservation is structural, while learning is devoted to task-directed routing. The same formulation identifies an identity-shortcut ResNet with its residual output retained as a full-rate PR system. Experiments verify exact single-channel routing of linearly separable factors to machine precision. On TIMIT, the proposed front-end improves test PER from $28.60\pm2.09\%$ to $25.76\pm0.41\%$ with the recognizer and training protocol held fixed, while maintaining exact reconstruction. Speaker probing further shows that structural conservation does not itself imply task-specific invariance.

Abstract: Extending large-scale multilingual automatic speech recognition (ASR) models to low-resource languages remains challenging. Model performance is skewed toward high-resource languages and degrades sharply for languages with limited labeled data and pre-training exposure. To address this, we investigate parameter-efficient approaches for transferring knowledge from resource-rich source languages to low-resource target languages on Whisper. Alongside warm initialization and attention-based fusion, we propose Sequential Adapter Stacking, which places a trainable target-language adapter on top of a frozen source-language adapter. Under controlled experiments, these approaches are evaluated on three target languages unsupported by Whisper -- Asturian, Assamese, and Xhosa -- using source languages with varying degrees of relatedness. Sequential Adapter Stacking with the closest related source consistently and significantly outperforms full fine-tuning across the three targets, with 5--8\% relative WER reductions. These gains largely persist with only one hour of target training data.

Abstract: For our submission to the WMT26 Creole Language Translation Shared Task, we focus on machine translation (MT) models for Pacific creoles: Tok Pisin, Bislama, and Solomon Pijin, with particular attention to broad domain performance. After pre-training on a large collection of domain-imbalanced data, we continue fine-tuning on a diverse mix of domain-balanced data. We rely on a number of data collection and preparation techniques, including LLM-assisted respelling and alignment, back-translation, and distillation from Gemini for domains originally not present in training data. Evaluated on Bouquet and a novel test set made of spoken language transcripts, our models beat open model baselines by 3+ chrF++ points in all directions with human-original references. Looking ahead, we plan to develop human-translated test sets for Solomon Pijin and Bislama, and to distil our best models into much smaller ones that retain broad domain coverage.

Abstract: We report a multi-month engineering program to build Sophea, a production bilingual Greek-English automatic speech recognition system. We evaluate the system against nine production gates covering Greek and English word error rate, language identification, and hallucinations on non-speech audio. Across twenty-three training iterations and two model architectures, no training-data composition passed all nine gates simultaneously. Meeting the Greek noisy-environment target required about 1,500 steps of dense domain exposure, while preserving English language identification tolerated only about 250 steps, or about 1,250 with a rebalanced mix that reduced Greek accuracy. We describe a six-stage data pipeline in which calibrating an audio-quality filter against in-domain anchors reduced the discarded share of scored Greek audio from 98.7 percent to 10.6 percent. A pre-registered ablation isolated a hallucination defect to one training-data package. A three-model ROVER ensemble increased gate coverage from 4-7 of 9 for individual models to 9 of 9 and reduced overlapping-speech WER from 53.35 percent to 37.87 percent, a 29 percent relative improvement. A separate learned per-clip arbiter over two models is listed as sophea/asr-k1 (preview) on the public Open ASR Leaderboard, with 4.26 percent average WER across eight public English test sets, and reaches 25.88 percent WER on live Greek noisy-environment traffic. We also document five cases in which a measurement tool produced a plausible but incorrect result and seven substantial approaches that were evaluated but not shipped. No model weights or training data are released; we report methodology and quantitative results only.

Abstract: Text-aligned speech tokenization methods have emerged to better align speech tokens with LLM token spaces, enabling more effective utilization of pretrained LLMs. However, they rely on offline automatic speech recognition (ASR), leading to two key limitations: (i) the need for complete utterances before tokenization, precluding real-time streaming, and (ii) vocabulary mismatch between ASR and LLMs, which reduces acoustic granularity from the subword to the word level. We introduce StreamAlign, a text-aligned speech tokenization framework that enables streaming tokenization for real-time speech-text joint modeling. StreamAlign performs online speech-text alignment by combining character-level RNN-Transducer alignment with word-level ASR guidance, mitigating ASR-LLM vocabulary mismatch while preserving recognition accuracy. A proactive word boundary classifier anticipates word completion at chunk boundaries, reducing tokenization latency from 560 ms to 270 ms. On LibriSpeech, StreamAlign achieves the lowest WER and highest UTMOS among evaluated tokenizers. Furthermore, StreamAlign-SLM, a spoken language model trained on StreamAlign units, outperforms other end-to-end spoken language models in speech continuation while achieving the strongest overall consistency on SALMon and spoken StoryCloze.

Abstract: This paper proposes a general framework for stable and effective iterative audio separation with mixture consistency by extending source separation models to a multi-input multi-output (MIMO) configuration. In the field of audio separation, mixture consistency is an essential property for many applications that require accurate phase and timbral information of target sources. While iterative approaches such as diffusion models achieve perceptually superior results in speech enhancement or user-guided target source separation tasks, most existing methods focus on single-step separation with a single-input single-output (SISO) or single-input multi-output (SIMO) configuration through architectural improvements, since mixture-consistent audio separation is generally regarded as a regression problem that admits a unique solution. By extending these architectures to a MIMO configuration, we introduce iterative prediction without compromising the architectural advantages or the characteristics of mixture consistency. We conduct a comprehensive ablation study of combining the framework with discriminators and extending it to a generative model. Experimental results demonstrate significant performance improvements when applying the proposed framework to state-of-the-art separation models.