speech technology teams

Abstract: Alzheimer's Disease (AD) detection from speech benefits from both paralinguistic and linguistic cues, yet model training is frequently constrained by limited data. Fine-tuning a large pretrained speech encoder can easily overfit, while a frozen probe leaves the encoder's task-relevant knowledge underutilized. Therefore, we propose Dual Task Vector Merging, a parameter-efficient method that adapts a frozen wav2vec 2.0 backbone using task knowledge learned outside the AD task. Specifically, we inject two task vectors, derived as weight deltas from emotion- and ASR-fine-tuned wav2vec 2.0 models, into each transformer layer. During AD training, only layer-wise merging coefficients and a linear classification head are optimized, updating less than 0.1% of the parameters required for full fine-tuning. Speaker-independent experiments on ADReSS and ADReSSo show that dual merging improves over frozen probes, LoRA, and full fine-tuning, in the main binary settings, and gives the best results when combined with LoRA. Evaluation on PROCESS further shows that our advantages transfer to three-class HC/MCI/dementia classification when paired with LoRA. Experiments under reduced training data (20%, 40%, and 80%) confirm reliable gains in low-resource settings.

Abstract: The SLAM-ASR framework connects the features of a frozen speech encoder as input to a frozen LLM decoder to enable speech transcription. In this work we ask whether this training-on speech transcription-provides the LLM with the capability to also perform spoken language understanding (SLU) tasks, by training monolingual projectors on 11 languages. When prompted to answer a yes/no question about the input speech, the mean instruction following rate (IFR) is 49.8%. When not following the instruction, these models default back to transcription. The transcription quality-measured by the character error rate (CER)-is not sufficient to predict IFR. Remarkably, when the model is instructed to answer the topic question as a single digit option, the mean IFR rises to 81.3%. Our models score a mean 23.2% topic balanced accuracy (BAcc) against 14.3% chance. However, the same decoder scores a mean 71.5% BAcc when using reference transcripts as text input. This indicates a failure of the projector to transfer information for the SLU tasks.

Abstract: We show that continuous bitstream diffusion (CoBit), originally developed for language modeling, provides a unified framework for joint and conditional multimodal generation using the same training and sampling protocol. Across image-text, speech-text and protein sequence-structure modeling, models are trained separately from scratch, but each uses a single input projection, transformer backbone, bitwise prediction head and binary score-matching objective shared across its modalities. Within each setting, a single checkpoint supports joint generation and the conditional directions included in training by changing which parts of the bitstream are observed. We first validate the approach on MNIST-Sum, a controlled image-text task pairing raw binary images with equations encoded as text, achieving near-perfect conditional accuracy and high joint consistency without learned tokenizers. We then apply the same protocol at larger scale to speech-text and protein sequence-structure modeling. Trained without text pretraining, a single speech-text model performs speech synthesis, recognition, continuation, and unconditional co-generation. Jointly generated speech and text achieve 3.5% word error rate between the generated text and automatic transcriptions of the generated speech. In protein modeling, a single CoBit checkpoint supports forward folding, inverse folding, and joint sequence-structure generation. For jointly generated proteins spanning 100-500 residues, computational refolding of CoBit's sequences yields structures that agree with the generated backbones at a mean self-consistency TM-score of 0.911. Together, these results show that CoBit provides a common generative framework for substantially different modalities and forms of cross-modal dependence, without introducing a domain-specific objective or sampling procedure.

Abstract: Data augmentation is standard practice for low-resource automatic speech recognition (ASR), but its benefits are almost always reported as aggregate error rates, leaving open why and when a method helps. On four low-resource languages and English at five data sizes, we cross two augmentations that modify existing audio (speed perturbation and SpecAugment) with two that modify text transcriptions and synthesize audio from it. We analyze errors at the phoneme level and find that they help for complementary reasons. SpecAugment has the most reliable aggregate improvements, but increases tone and palatalization errors, whereas text-based methods introduce consonant insertions; in combination, each reduces the errors the other introduces. This complementarity is predictable. Tone and palatalization are each signaled by a single acoustic property and occur least frequently in training (lowest training exposure), and text-based methods that flatten word frequency raise their frequency by 37--38%, reversing those errors while further improving aggregate error. The improvement from combining the methods exceeds the sum of their individual improvements; the superadditivity vanishes by 100 hours, limiting the complementarity to the low-resource regime. Where specific error types matter, we recommend selecting augmentation by training exposure and acoustic properties, both measurable before any training.

Abstract: Speech large language models (speech LLMs) perform well on automatic speech recognition (ASR) when sufficient paired speech-text data is available, but their performance degrades in low-resource settings. A cascaded pipeline that performs speech-to-phoneme (S2P) conversion followed by phoneme-to-grapheme (P2G) conversion has been shown to outperform end-to-end speech LLMs in this regime, suggesting that phoneme-mediated processing is beneficial when paired data is scarce. We propose \textit{phoneme-guided initialization}, a simple method that uses this insight within an end-to-end framework: we pre-train the audio encoder on S2P and the LLM on P2G tasks, then connect them and fine-tune the full model end-to-end on the target ASR task. Experiments on Japanese (CSJ), Chinese (AISHELL-1), and two low-resource languages from Common Voice 25.0 (Tatar and Urdu) show that our method matches or outperforms both the cascaded S2P-P2G baseline and the end-to-end model without P2G initialization.

Abstract: Audio language models must handle dozens of distinct skills, from pitch comparison and speaker counting to musical tempo estimation and emotion recognition. Joint training on all skills at once causes interference: gains on one skill often come at the cost of another. We propose \textbf{SkillFormer}, which decomposes audio understanding into skill-specific low-rank adapters and composes them at inference time through a learned router. The router examines the question to decide which adapters to activate and how much weight each should carry, so that a pitch query engages different parameters than a genre classification query. An alternating training schedule updates each adapter on its own skill cluster before jointly calibrating the router, preventing the gradient conflicts that arise in standard multi-task optimization. SkillFormer adds fewer than 4\% of the base model's parameters and requires no changes to the audio encoder or language backbone. Evaluated on three architecturally distinct models across MMSU, MMAU-Pro, and MMAR, it raises the average accuracy by 2.5 to 4.1 points, with balanced gains across perception, reasoning, and semantic subcategories.

Abstract: Although reinforcement learning (RL) post-training repairs the localized segmental errors of zero-shot text-to-speech (TTS), arriving at a working recipe still relies on tedious manual tuning, and whether LLM agents can take over this research pipeline is unclear. We investigate this question with AgenticTTS-Forge, a collaborative workflow that structures human guidance and agentic execution around a shared workspace, applied to CosyVoice2-0.5B. To measure what the agent automates, we audit its trajectory stage by stage against the published recipe. To measure what it exploits, we score its policies with held-out observers hidden from the agent. Our results show that the agent recovers an underspecified recipe, improves it, and, when gains stall, surveys the literature unprompted and pivots from the LM carrier to the flow carrier, halving Bad cases. However, its autonomy exposes three traps across the data, proxy, and algorithm axes: the held-out set leaks through a channel the contract never reads, a self-shaped reward inflates the proxy where it is scored, and separately tuned policies do not compose additively. These findings show that the binding constraint is measurement rather than reasoning, and can inform the design of harnesses whose contracts read every channel the agent does.

Abstract: Speech anti-spoofing countermeasures degrade when the generator, codec or channel changes, and a common remedy is an auxiliary objective that shapes the embedding space; whether it does is invisible to EER, a pure ranking metric. We compare seven such objectives with cross-entropy over 113 runs on five corpora, AASIST3 at three seeds plus four pre-trained detectors, and measure the embedding space of the 24 AASIST3 runs directly. Raw augmentation displacement makes cosine consistency look effective, but the gain is a smaller space, not a more stable one: normalised by the spread, no configuration consistently improves on cross-entropy. Every trained space is dominated by the single decision axis expected for two classes, whose training-set structure does not transfer, and four runs collapse to a near-constant output that displacement rewards and EER reports as poor accuracy. No auxiliary objective keeps an advantage over cross-entropy across architectures, corpora and seeds.

Abstract: Zero-shot text-to-speech synthesizes new utterances in a speaker's voice from a short reference recording. Voice cloning requires accurate content and preserved speaker identity, but supervised acoustic-token prediction does not directly optimize these waveform-level properties. Reward-based post-training addresses this mismatch, but in discrete diffusion, token choices and reveal positions jointly define the sampling trajectory, complicating alignment. We introduce RAWD-TTS (Ratio-free Advantage-Weighted Denoising), which scores decoded samples with recognition and speaker rewards and uses group-relative advantages to weight masked-token reconstruction of those samples, without reverse-trajectory likelihoods or target audio. On 500 Russian CV3-Eval voice-cloning prompts, joint alignment reduces word error rate from 3.18% to 2.42% at the reward-selected checkpoint (24.0% relative) and to 2.58% at the final checkpoint (19.0%), while WavLM speaker cosine rises from 0.733 to 0.748 and 0.755. Controlled experiments characterize recognition-identity trade-offs and the effects of corruption count, group composition, and weighting.

Abstract: Multilingual self-supervised speech models can benefit from sharing information across languages, but under a matched total pretraining data budget they still fall short of monolingual models. We show that strengthening the model's ability to discriminate languages during pretraining reduces and, on some measures, closes this multilingual gap on continuous phonetic and higher-level linguistic measures, while preserving substantial cross-language sharing. Using a controlled English/French HuBERT setting, we test two interventions which strengthen language discrimination: an auxiliary language classifier and per-language k-means targets. Across interventions, continuous-feature phone discrimination error (phone-ABX, lower is better) decreases from 11.6% in the bilingual baseline to 10.4% (monolingual: 10.8%), while lexical performance (sWUGGY, higher is better) increases from 52.1% to 56.7% (monolingual: 58.5%) and prosodic performance (ProsAudit, lexical subtask, higher is better) from 68.9% to 72.9% (monolingual: 72.6%). Across HuBERT training stages, the strongest gains on most linguistic measures occur when language discrimination is introduced in the first iteration, whereas later or repeated interventions yield smaller improvements and are accompanied by increased language-wise segregation. These results support a causal role for language discrimination in reducing the additional cost of multilingual learning.

Abstract: Audio language models understand what is said far better than how it sounds. Closing this gap takes more than data. Detailed acoustic annotation is costly, labels from stronger models inherit their errors and limits, and fixed data cannot adapt as the learner improves. We therefore propose EvoAudio, a recursive self-improvement system for audio understanding. To our knowledge, it is the first to evolve the model, waveforms, questions, and difficulty in one closed loop. EvoAudio uses the current model's performance to set the focus and difficulty of the next training data. A library of audio tools then constructs questions whose answers follow from how the audio was made, providing verifiable supervision without new human annotation. Reinforcement learning updates the model, and validation decides whether it enters the next evolution round. Across 13 rounds, EvoAudio improves five models with different audio encoders and language backbones on MMSU, MMAU-Pro, and MMAR. It achieves the highest average for every backbone, raising overall performance by up to 6.3 points. The improvement unfolds over successive rounds, with each stronger model starting the next round.

Abstract: Compositionality, the ability to represent complex acoustic scenes as combinations of simpler sound sources, is central to auditory perception and classical additive signal models. Still, it remains unclear whether modern pre-trained audio representations internalize additive structure without compositional supervision. Existing evaluation frameworks of audio compositional reasoning largely focus on cross-modal audio-text alignment, leaving open whether audio representations themselves exhibit additive compositional structure independent of text grounding, analogous to vector arithmetic in word representations. To investigate this, we adopt a two-step diagnostic for frozen audio representations. First, we quantify linear alignment between representations and sound source labels using canonical correlation analysis. Second, we test additive compositional generalization via leave-one-combination-out reconstruction, grouping clips by exact source-label set, averaging their representations, and predicting held-out means from per-source contributions fitted only on training combinations. With larger combination holdouts, CLAP outperforms the permuted and label-overlap baselines on FSD50K, while the speech models do not outperform the label-overlap baseline. We examine representations generated by Wav2Vec2, HuBERT, and CLAP on FSD50K and CHiME-Home datasets. All three models show consistently higher linear correlation and more accurate leave-one-combination-out reconstructions than the permuted baselines. However, only CLAP shows large cosine similarity gains, which could be associated with its training on many kinds of audio and text. Finally, we note that all three models exhibit reconstruction residuals, revealing limits of additive compositionality such as nonlinear or non-compositional audio structure.

Abstract: Evaluating live streaming speech synthesis (TTS) requires assessing fine-grained, highly expressive prosody such as emotion, intonation, and energy which traditional MOS predictors fail to capture. While proprietary Large Language Models (LLMs) like Gemini can evaluate these aspects, they are too costly for massive inference and reinforcement learning feedback. To address this, we first introduce Live-ProsodyJudge (LPJ), a cost-effective pairwise evaluator distilled from Gemini into Qwen3-Omni. However, we identify a critical flaw in standard multi-dimensional evaluation: verdict coupling. The judge tends to lazily align all individual dimension scores with its overall preference, collapsing a rich multi-dimensional rubric into a single preference bit. To resolve this, we further propose Decoupled-Live-ProsodyJudge (D-LPJ). D-LPJ eliminates the overall verdict target to prevent blind following, masks uncertain pair-dimensions during Supervised Fine-Tuning(SFT), and introduces a novel span-local GRPO strategy that applies normalized advantages strictly to their corresponding rationale spans. Evaluated on highly curated human-annotated test sets, 10 sample balanced-order LPJ achieves higher point accuracy than a single Gemini call, while D-LPJ successfully produces independent,decoupled dimension judgments. Furthermore, in a Best-of-8 TTS candidate selection tournament, the LPJ-selected utterance falls within the human top-3 in 85.29% of high-confidence cases, demonstrating its efficacy for fine-grained TTS preference optimization.

Abstract: Training speaker diarization models on loose labels, such as speech segments with padded boundaries or filled pauses, often results in similarly loose model outputs. To obtain tighter boundaries, pseudo-labeling based on the averaged outputs of causal and anticausal models has been proposed. However, since the pseudo-labels are estimation-based, they can suffer from over-tightening, which increases missed detections that can propagate as unrecoverable errors to downstream tasks. This paper carefully analyzes the causes of over-tightening and proposes three approaches to address them: (i) removing pause filling rather than padding, (ii) introducing a burn-in phase to mitigate missed detections near the beginning of causal and anticausal predictions, and (iii) making pseudo-label-based co-training aware of the non-causal model used for final inference. Experimental results show that the proposed method reduces missed detections caused by over-tightening and improves both diarization accuracy and downstream multi-talker ASR performance.

Abstract: Simultaneous speech-to-speech translation requires understanding, translation and spoken delivery while the source stream continues. To support timely delivery and limit accumulated delay, systems adopt reformulation and summarization, which can preserve meaning while departing from written references. BLEU and COMET may not reliably distinguish such variation from semantic loss. We introduce EviSI, a large language model evaluation agent adapting the error analysis and penalty principles of Multidimensional Quality Metrics (MQM). It constructs shared source evidence, assesses semantic fidelity and oral expression, reconciles overlapping errors and scores deterministically. EviSI recovers the aggregate human system ranking for English to Chinese. Mean Kendall agreement with human system rankings within corpora reaches 0.707 for English to Chinese and 0.467 for Chinese to English, exceeding evaluated baselines. An extension across five directions shows positive concordance with COMET without human ratings. Individual output agreement with humans remains mixed.