Speech Language Models

Ask about this topic

Overview

Speech language models connect spoken audio to language-model capabilities, supporting tasks such as transcription, spoken question answering, translation, diarization, and speech generation. Common approaches pair pretrained audio encoders or speech tokenizers with language models, using projectors, alignment objectives, interleaved speech-text training, or distillation to bridge the modalities. Post-training and decoding methods extend these systems to streaming and full-duplex conversation, where turn-taking, latency, and expressive speech matter alongside answer quality. Models are becoming more capable across tasks, but robustness to noise, overlap, low-resource languages, and acoustic cues remains uneven, and end-to-end systems do not always outperform cascades of specialized components.

Latest papers

CardsList

312 papers55 papers in the last four weeks, up 8% on the four weeks before. 0.3% of all new papers.

Sep 11, 2026

Causal Analysis and Mitigation of Spurious Onsets in Full-Duplex Speech LLMs

Speech-to-speech LLMs like Moshi, and its derivative PersonaPlex, can listen and speak concurrently through full-duplex generation. However, they can begin speaking inappropriately during prolonged user silence: under digital-zero input, Moshi and PersonaPlex initiate speech in 30% and 27.5% of five-minute continuations, respectively. What causes this spurious speech? We investigate two hypotheses: either repeated sampling selects speech despite persistently low onset probabilities, or self-conditioning on nonspeech outputs causes an abrupt spike in onset probability. We find that, at every observed onset, speech probability spikes by over nine orders of magnitude in one 80-ms frame, supporting the latter hypothesis. Then, to suppress these onsets without blocking genuine responses, we ask a causal counterfactual question: is the model responding to user speech, or would its next-token distribution remain similar if the preceding user input were muted? Accordingly, we suppress onsets whose distributions change little under this intervention. Under realistic microphone noise, our method suppresses spurious onsets, while preserving genuine responses: one-sided 95% lower confidence bounds are 98.68% and 98.82% for Moshi, and 96.90% and 99.25% for PersonaPlex. Our inference-time method runs in real-time without retraining, with 95th-percentile decision time below 61 ms, within the 80-ms frame budget. Our code is available at https://github.com/KentoNishi/icassp27-spurious-onsets.

Kento Nishi

Massachusetts Institute of Technology · Comcast Applied AI Research, Speech AI Team

Speech Language ModelsCausal Interventions in Language ModelsFull-Duplex Spoken Dialogue Systems
AskPDF
Sep 11, 2026

Kraken: LLM-based Speech-to-Speech Translation via Low-bitrate VQ and Dual-path Source Conditioning

Speech-to-speech translation (S2ST) has advanced significantly with speech LLMs, offering the potential for joint optimization and preserving non-linguistic information. However, these models struggle with predicting high-bitrate speech tokens in LLMs, and face the challenge of relying on S2ST training data with ideally aligned speaker identity and prosody. We propose using low-bitrate tokens based on single-layer vector quantization, trained to reconstruct self-supervised learning (SSL) features. We also employ a separate token-to-waveform decoder named Autowave-X, which is also conditioned on the source speech to improve non-linguistic transfer, thereby relaxing the training data constraints. With the integration of these techniques, we propose an S2ST model named Kraken, which augments a pre-trained LLM with speech feature inputs and the low-bitrate token outputs, followed by Autowave-X vocoder. We built the model upon Qwen3-8B and trained it using 150k hours of multilingual and multitask speech data. We demonstrated that our model exhibited better translation quality than SeamlessM4T-Large v2 and Qwen2.5-Omni, along with improved speaker and prosody transfer capabilities.

Hayato Futami, Hassan Shahmohammadi, Tushar Dhyani +5

Sony Group Corporation · Sony Europe Limited

Speech TranslationSpeech Language ModelsDiscrete Speech Representation Learning
AskPDF
Sep 10, 2026

Nuha-Speech: Building General-Purpose Arabic Speech-LLMs

As Speech Large Language Models (speech-LLMs) become increasingly multilingual, Arabic remains significantly underrepresented, highlighting the need for dedicated infrastructure to train and evaluate Arabic speech-LLMs. To address this gap, we introduce Nuha-Speech, a comprehensive initiative to develop general-purpose Arabic speech-LLMs spanning dataset construction, model training, and systematic evaluation. Specifically, we constructed a large-scale Arabic Speech Question-Answering (SQA) corpus comprising over 1.5 million training samples to allow instruction tuning over a broad range of core speech tasks. Then, the corpus was used for supervised fine-tuning based on Qwen-Omni model variants at different scales. Finally, we designed an evaluation framework featuring diverse tasks and tailored metrics. Through this work, we aim to establish foundational infrastructures for Arabic Speech-LLMs under constraints imposed by limited Arabic speech resources.

Yingzhi Wang, Reem Alhazzani, Muhammad Alqurishi

Elm Company, KSA

Audio-Language Model EvaluationAudio QAArabic NLPArabic Language ModelsSpeech Language Models
AskPDF
Sep 10, 2026

RetroThinker: Enabling Retrospective Thinking in Speech LLMs

Speech large language models (SpeechLLMs) offer reduced latency and retain paralinguistic nuances that are typically lost in cascaded automatic speech recognition (ASR) and text-based LM architectures. However, they continue to lag behind text-only LLMs on complex reasoning tasks, while real-time spoken interaction imposes strict latency constraints. Although prior works employ Chain-of-Thought (CoT) and concurrent reasoning to enhance reasoning capabilities without inducing prohibitive delays, an inherent accuracy-latency trade-off persists. In this paper, we investigate whether a streaming SpeechLLM can dynamically revise its reasoning traces on the fly. We introduce RetroThinker, a multi-stage post-training framework that equips the Moshi model to self-verify and forward-correct CoT steps during inference. RetroThinker combines supervised fine-tuning (SFT) on curated retrospective thinking data with length-based direct preference optimization (DPO) to optimize retrospective during early reasoning (i.e., reasoning concurrently while the user speaks). Evaluated on the GSM8K benchmark, RetroThinker significantly improves the accuracy-latency trade-off over non-retrospective baselines, achieving an 11% absolute accuracy gain at a comparable latency.

Yi-Jen Shih, Puyuan Peng, Abdelrahman Mohamed +1

The University of Texas at Austin · FAIR, Meta Superintelligence Labs

LLM Self-CorrectionEfficient Language Model ReasoningSpeech Language Models
AskPDF
Sep 10, 2026

Component-Aware Differential Privacy for Federated Multilingual Speech-LLMs

Per-layer differential privacy (DP) clipping improves gradient fidelity in federated learning by allocating per-matrix clipping budgets proportional to parameter count. We show that this recipe breaks for speech large language models (speech-LLMs), when the acoustic encoder and the language decoder differ by an order of magnitude in update norm. Single-pool per-layer methods suffer cross-component budget collapse, dragging word error rate (WER) far from flat global clipping or collapsing training entirely. When the norm imbalance is milder, adaptive single-pool methods partially recover, confirming that collapse severity scales with the inter-component norm ratio. We empirically diagnose the root cause across six per-layer methods and three speech-LLM architectures. We then propose \emph{αα-split}, a two-pool allocation that normalises encoder and LLM parameters into independent pools, and show that joint ℓ2\ell_2 sensitivity and the original (ε,δ)(\varepsilon,δ)-DP guarantee are unchanged. At architecture-calibrated αα, our method recovers WER utility compared to flat DP, while granting the encoder 4.47×4.47{\times} tighter per-component noise protection against speaker voice-based gradient-inversion attacks at only +2.6%+2.6\% LLM noise overhead.

Jordi Luque, Fernando López, Aleix Sant

Scientific Research Telef´onica Innovaci´on Digital Barcelona, Spain · Telef´onica Innovaci´on Digital Universidad Aut´onoma de Madrid Madrid, Spain · Scientific Research Telef´onica Innovaci´on Digital Universitat Polit`ecnica de Catalunya Barcelona, Spain

Privacy-Preserving Language ModelsDifferentially Private Federated LearningDifferential PrivacySpeech Language ModelsGaussian Differential Privacy
AskPDF
Sep 10, 2026

The Eloquence submission for Task 2 of the Interspeech 2026 MLC-SLM challenge

This paper details the Eloquence team's approach to Task 2 of the 2nd MLC-SLM challenge at Interspeech 2026, which involves multilingual Multiple-Choice Question Answering (MCQA) across 21 languages. Three approaches are explored. First, we fine-tune Voxtral-Mini-3B via LoRA with cross-lingual data augmentation, ASR transcript augmentation and timestamp-aware audio cropping, achieving 0.72 macro-accuracy on evaluation Phase 2. Second, we apply multimodal in-context learning (ICL) to the frozen Voxtral-24B model to correct a strong label bias, reaching 0.81, our best result. Third, a training-free retrieval system based on a three-layer voice-anchored memory combining acoustic identity, semantic content, and a knowledge graph achieves 0.68. All three systems substantially outperform the official baseline.

Jordi Luque, Lorenzo Concina, Marco Matassoni +2

Telefónica Innovación Digital, Scientific Group · Fondazione Bruno Kessler · Consiglio Nazionale delle Ricerche

Multilingual Language Model EvaluationMultilingual QASpeech Language Models
AskPDF
Sep 10, 2026

X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation

Reducing audio-encoder depth lowers the inference cost of speech large language models, but removing complete blocks perturbs the embeddings consumed by the decoder and can cause deletion and premature end-of-sequence errors. We introduce X-AuT, a progressive framework that selects layer combinations through short behavioral probes and restores the pruned model through representation alignment, cross-scale distillation, scheduled student-policy supervision, and LoRA finetuning. The language-model backbone remains frozen, while attention LoRA adapters and the tied output embedding adapt during distillation. Training uses the highest-agreement tier from a transcript-consistency pipeline, followed by source reweighting during finetuning. On ten public Chinese--English benchmarks, compressing Qwen3-ASR-0.6B from 18 to 16 audio-encoder layers reduces macro-average error from 5.61% to 5.27%. The 14-layer model reaches 5.75% with 20.7% fewer audio-tower parameters. Under the matched recipe, the 1.7B teacher yields 5.55% mean error, compared with 8.45% for self-distillation, and progressive 18→\rightarrow14 pruning outperforms direct pruning (5.75% vs. 6.73%). These single-run results establish two practical operating points and show that the accuracy effects vary across benchmarks. Project website: https://xpeng-ai.github.io/x-aut

Haojun Zhang, Yi Zou, Min Chen +7

XPeng Inc.

LLM CompressionLanguage Model DistillationSpeech Language ModelsNeural Network Pruning
AskPDF
Sep 10, 2026

SEAR: Segment-Evidence-Aware Routing for Weak-to-Strong Multilingual Speech MCQ

This paper describes our system for Task2 of the second Multilingual Conversational Speech Language Model (MLC-SLM) Challenge. We adapt Qwen3-Omni-30B-A3B-Instruct with a segment-evidence-aware data and post-training pipeline. A language model converts timestamped ASR into coherent event spans, which are expanded by a boundary margin and cropped from the original recording. We then synthesize complementary semantic MCQs with Qwen3.6-27B and acoustic MCQs with Gemini3.1 Flash-Lite, followed by structural, grounding, answer-consistency, and target-model trainability checks, yielding 359,825 verified MCQs across 21 language and accent variants. A text-only probe partitions the data into weak, text-answerable items used for supervised fine-tuning and strong, audio-dependent items used for reinforcement learning with Group Sequence Policy Optimization (GSPO), stabilized by debiased advantages, sequence-level importance correction, and dynamic filtering. Our system obtains 90.92% accuracy on the final official evaluation set.

Huy Hoang Le, Long-Bao Nguyen, Minh Tri Dao

CAKE by VPBank, Vietnam

Audio-Language Model EvaluationSpoken Language UnderstandingMultilingual QASpeech Language Models
AskPDF
Sep 10, 2026

Xiaomi-CocktailASR-1 Technical Report

Recently, large language model (LLM) based ASR models have achieved significant progress, yet they generally lack support for multi-speaker scenarios, where the cocktail party problem remains a critical bottleneck for further advancing ASR. Existing TS-ASR methods, including end-to-end architectures with speaker embeddings and latest LLM-based explorations suffer from degraded single-speaker performance and the inability to reject when the target speaker is absent. In this paper, we propose Xiaomi-CocktailASR-1, an LLM-based end-to-end TS-ASR architecture. By utilizing reference speech as voiceprint prompts, it directly transcribes the target speaker's speech without requiring speech separation. Xiaomi-CocktailASR-1 maintains competitive performance in single-speaker scenarios, comparable to mainstream ASR models. It also features a negative sample rejection capability, outputting empty text when the target speaker is absent from the mixed speech. Additionally, Xiaomi-CocktailASR-1 supports a Chain-of-Thought (CoT) reasoning mode to provide explicit reasoning steps. Extensive experiments on various synthetic and real-world multispeaker benchmarks demonstrate that Xiaomi-CocktailASR-1 achieves state-of-the-art performance, effectively addressing the cocktail party problem through a unified architecture that balances multispeaker and single-speaker recognition accuracy, along with rejection capability.

Yiru Zhang, Hang Su, Lichun Fan +10

Xiaoai Plus, ASR team Xiaomi Inc., China

Automatic Speech RecognitionSpeech Language Models
AskPDF
Sep 9, 2026

LLM-Anchored Paralinguistic Enrichment for Alzheimer's Disease Detection

Speech-based automatic detection of Alzheimer's disease (AD) provides a non-invasive and scalable approach to early cognitive screening. AD affects both lexical-semantic organization and speech production, including atypical pauses and word elongations. However, existing methods have yet to fully integrate these paralinguistic cues with linguistic content. We propose LLM-Anchored Paralinguistic Enrichment (LAPE), which enriches LLM-derived linguistic representations with paralinguistic cues through three coordinated innovations. The first is prosodic event textualization, which enables the LLM to model pauses and elongations jointly with lexical content by encoding them as explicit markers with bounded duration-aware repetition. The second is lexico-prosodic unitization and chunking, which preserves event identity and magnitude in both modalities by pooling only consecutive word units. The third is text-anchored paralinguistic fusion, which integrates local and utterance-level speech features by using NormGate to normalize and dynamically scale them relative to text. We evaluate LAPE on ADReSS and ADReSSo using participant-level cross-validation and leave-one-subject-out evaluation. LAPE achieves state-of-the-art performance across all four primary settings. Code will be released upon acceptance.

Xiao Wei, Yuqin Lin, Yaru Cao +6

Tianjin Key Laboratory of Cognitive Computing and Application, College of Intelligence and Computing, Tianjin University, Tianjin, China · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China · College of Computer and Data Science, Fuzhou University, Fuzhou, China +1

Speech-Based Dementia DetectionSpeech ProsodyAlzheimer's Disease ClassificationSpeech Language Models
AskPDF
Sep 9, 2026

StreamAlign: Streaming Text-Aligned Speech Tokenization

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.

Kang-wook Kim, Jinyoung Park, Jinsoo Kim +3

Seoul National University · University of California, Berkeley · KRAFTON +1

Audio TokenizationSpeech Language ModelsDiscrete Speech Representation Learning
AskPDF
Sep 8, 2026

TASTE2: Text-Aligned Speech Modeling and Deployment toward Full-Duplex Voice Interaction

Full-duplex voice interaction requires more than utterance-level conversion. It must process streaming speech, manage turn-taking and interruptions, while preserving pretrained linguistic competence and acoustic paralinguistic cues. We ask whether TASTE (Text-Aligned Speech Tokenization and Embedding) provides a viable path toward this goal. We present TASTE2, which transforms utterance-level TASTE into an incremental dialogue stack. A shared text-token vocabulary removes word-level averaging, while modality-aligned dialogue training predicts one continuous audio latent per text token without interleaving heterogeneous token streams. An incremental Speech Detokenizer enables streaming synthesis through CosyVoice2. After speech and dialogue training, TASTE2 (Merge) reaches 56.3% on LLaMA-Questions against a 57.3% Qwen2.5-7B Instruct text-only reference (98.2% accuracy retention), and TASTE2 (Direct) reaches 53.0% (92.4% retention). We build TASTE2 VoiceBot, which processes user speech incrementally, streams synthesized audio, and stops generation on barge-in. On Full-Duplex-Bench v1.0, TASTE2 and TASTE2 VoiceBot handle interruptions well while maintaining high conversational coherence. Natural conversation remains challenging, and deployed mean time to first audio is 2.701 s on two NVIDIA RTX A6000 after TensorRT acceleration. Finally, to our knowledge, we provide the first systematic characterization of explicit paralinguistic control in a TASTE based model. Fast speaking rate serves as a cross-strategy proof of concept after dialogue SFT, while emotion control is strategy dependent and the remaining attributes stay weak. Together, these results establish TASTE based modeling as a practical route toward full-duplex systems while identifying natural conversation robustness, speech generation latency, and feature general paralinguistic control as open challenges. Explore TASTE2 online.

Yi-Chang Chen, Chun Wei Chen, Dien-Ruei Wu +7

MediaTek Research · Internship at MediaTek Research · National Taiwan University +2

TTS SynthesisStreaming TTS SynthesisControllable Speech GenerationSpeech Language ModelsFull-Duplex Spoken Dialogue Systems
AskPDF
Sep 8, 2026

TontaubeV1: Streaming Text-to-Speech with Hierarchical Codec Modeling and Bounded Context

Text-to-speech systems often face a trade-off between natural prosody and efficient inference: higher perceptual quality usually comes at increased computational cost and latency. We present TontaubeV1, a model that preserves natural prosody while enabling streaming from a single consumer GPU. Speech is encoded by the hierarchical DualCodec representation at 12.5 Hz, which separates a semantic stream from successive acoustic refinements. Our design assumes that prosodic structure is largely established when the semantic stream is generated, and allocates capacity accordingly: a Qwen3-1.7B-derived transformer predicts that stream and thereby the utterance duration, while three progressively smaller Qwen3-0.6B-derived transformers each add one acoustic refinement. Text is tokenized per character rather than by subword. Paired text and audio markers at shared positions support long-form generation with bounded context, and overlapping DualCodec reconstructions are mapped into the VibeVoice acoustic latent space and decoded causally, enabling streaming despite DualCodec's noncausal decoder. The model accepts up to one minute of reference audio for voice conditioning and is designed primarily for English and German, with additional multilingual support. The four predictors total 2.9B parameters; on a single RTX 5090 the streaming path reaches about 200 ms to first audio. In separate non-streaming measurements, the end-to-end real-time factor (RTF) is 0.08 for one input and the aggregate RTF is 0.02 across eight concurrent inputs. In a blinded study, 40 listeners preferred TontaubeV1's prosody over Fish Audio S2 Pro in 70.25% of 400 comparisons (95% CI: 64.2-75.7%). On an LLM-as-a-judge benchmark, it matches ElevenLabs Flash v2.5 and outperforms the April 2026 Gradium API and Cartesia Sonic 3 on prosody. The weights are released on Hugging Face under the Tontaube Community Model License 1.0.

Fritz Cremer, Jonathan Cremer

TTS SynthesisStreaming TTS SynthesisLong-Form TTS SynthesisSpeech Language Models
AskPDF
Sep 4, 2026

TamilEOT: A Dataset and Model for Semantic End-of-Turn Detection in Tamil Telephone Speech

A voice agent has to decide, at every pause, whether the user has finished speaking. Without a model of the language that decision falls back to a fixed silence timeout: set it short and the agent interrupts, set it long and every turn pays the full wait. Open semantic end-of-turn detectors exist, but to our knowledge none covers a South Indian language. We release TamilEOT: 18,485 labelled turn boundaries cut from 116 real Tamil telephone conversations, and two audio-only detectors fine-tuned from Smart Turn v3. On a held-out split of 4,168 clips from 30 unseen calls, accuracy rises from 70.30% zero-shot to 83.71% (8.7 MB) and 86.13% (21 MB); ROC-AUC rises from 0.751 to 0.921. Both models run in under 150 ms single-threaded on a laptop CPU. We also report what building it cost. Rule-derived labels, checked against a blind human listening pass, were right 95.9% of the time on the positive class and 44.4% on the negative class, which is below chance, because the rule answered a different question than the model is asked. Replacing them with an audio-LLM labeller measured at 97.5% human agreement cost US$5.69. Of every training lever we measured, only encoder capacity moved the result; three runs at identical config and seed span 0.87 accuracy points, which is the floor below which none of our other deltas mean anything. Replaying the same labelled boundaries through the production VAD and streaming adapter costs a further 2.60 points, and 7.8% of boundaries are never surfaced to the model at all. Data, weights, code and every negative result are public.

Santhoshkumar V

Turn-Taking PredictionConversational AgentsSpeech Language Models
AskPDF
Sep 4, 2026

What Did I Just Say? Self-Listening for Full-Duplex Speech Models

Full-duplex spoken language models can listen and speak simultaneously, enabling them to handle interruptions and backchannels in human conversation. However, text generation, speech synthesis, and audio playback proceed asynchronously. As a result, what a model believes it has said may not match what has actually been played to the user. We refer to the problem of recovering from an interruption while remaining aware of the model's realized speech as anchor interruption. To address this problem, we propose Self-Listening, a full-duplex modeling approach that interleaves user speech, model text, and the model's played speech. By feeding the realized speech output back to the model as an input stream, self-listening grounds interruption recovery in what the user has actually heard. We further introduce AnchorSpeech, a collection with homogeneous training and test splits for tracking which items of structured ordered responses have actually been spoken. AnchorSpeech-test evaluates whether a model can respond consistently with the last completed item before an interruption. Experiments show that, compared with full-duplex baselines, models equipped with self-listening mechanism achieve better anchoring performance.

Xuanning Zhou, Junyi Ao, Xiaotong Liu +3

Shenzhen Loop Area Institute, China · The Chinese University of Hong Kong, Shenzhen, China

Speech Language ModelsFull-Duplex Spoken Dialogue SystemsSpoken Dialogue Systems
AskPDF
Sep 3, 2026

Decoupling Turn-Taking from Semantics: A Decoupled Data Approach for Finite-State-Machine-Based Full-Duplex Dialogue

The Neural Finite State Machine (NFSM) framework offers a pragmatic path to full-duplex dialogue by serializing turn-taking control and response generation onto a single causal tape under the standard next-token prediction objective, thereby preserving semantic prowess at a low fine-tuning cost. However, its reliance on synthetic text data fundamentally limits turn-taking naturalness, as Large Language Models (LLMs) cannot faithfully simulate the fine-grained acoustic temporal dynamics of real human dialogues. In this work, we propose a decoupled data approach that learns turn-taking from real Human-Human (HH) spoken dialogues while shaping semantic behavior through configurable Human-Agent (HA) text dialogues. To operationalize this approach, we introduce a rule-based event-guided data transformation method that serializes HH spoken dialogues into FSM tapes by classifying turn-taking events and applying deterministic mapping rules, enabling scalable supervision without LLM-generated annotations. We further propose a Source-Aware Calibrated (SAC) Loss that jointly calibrates the long-tailed distribution of state transition tokens and channels each data source toward the capability it best supervises. Experiments show that our approach substantially improves turn-taking proficiency while recovering the foundation LLM's semantic capability. Our code and model are available at https://github.com/Liyht/def-fsm.

Yihang Li, Chenhui Chu

Kyoto University

Spoken Dialogue SystemsSpeech Language ModelsTurn-Taking PredictionFull-Duplex Spoken Dialogue SystemsTurn-Taking in Dialogue
AskPDF
Aug 31, 2026

When Does Predictor-Based RL Align with Human Perception? A Study of Subjective Rewards in Codec-Based Speech Language Models

Codec-based text-to-speech (TTS) models make language-model post-training applicable to speech generation, but it remains unclear when learned perceptual predictors can serve as reinforcement learning rewards without losing alignment with human listeners. We study this question with Group Relative Policy Optimization (GRPO) using learned rewards for anime-like speaking style, naturalness, likability, and arousal. To prevent perceptual rewards from being optimized through transcript drift, we introduce a character error rate (CER) zone constraint and compare policy optimization with Best-of-NN reranking under the same reward gate. Across single-reward runs, each reward primarily improves its own target metric, showing that subjective predictors are not interchangeable quality surrogates. Multi-rater A/B tests further show uneven human transfer, while a reward-gap analysis separates average transfer from within-axis calibration: signed reward gaps significantly predict listener choices in the pooled analysis, whereas residual CER gaps do not, but per-axis calibration remains heterogeneous. Best-of-8 is a strong human-level baseline and is not clearly worse than GRPO perceptually, suggesting that GRPO should be viewed as amortizing reward-selected behavior into the policy rather than uniformly outperforming reranking. These results support analyzing subjective speech rewards as predictor-axis-base tuples and provide practical diagnostics for selecting rewards before multi-reward speech post-training.

Joonyong Park, Jerry Li

Spellbrush

Reward ModelingRL for Language ModelsSpeech Generation EvaluationTTS EvaluationSpeech Language Models
AskPDF
Aug 31, 2026

Listen to the Latents: Self-Correcting Speech Recognition in Large Audio Language Models Through Hidden-State Interactions

Recent automatic speech recognition (ASR) systems increasingly integrate large language models (LLMs) to leverage their semantic knowledge, either externally through logit fusion or internally through warm initialization. However, how to effectively combine these two strategies remains underexplored. In this work, we refine warm-initialized LLM-based ASR models by leveraging their own pre-adaptation base LLMs, focusing on LoRA-adapted settings where the base LLM is preserved. To achieve this, we propose Hybrid Search, a targeted correction strategy motivated by two observations. First, interaction features that characterize the relationship between LLM-based ASR hidden states and base-LLM hidden states provide informative signals about a token's degree of semantic dependence. Second, selectively refining targeted tokens with high semantic dependence improves ASR performance far beyond naive global LLM-correction methods including rescoring and late fusion. Our analysis suggests that, even after semantic knowledge transfer through warm initialization, LLM-based ASR models can still leverage their base LLM to further improve inference-time performance.

Chan-Jan Hsu, Jaeyeon Kim, Chao-Han Huck Yang +3

Carnegie Mellon University · NVIDIA · Work done at NVIDIA +1

Automatic Speech RecognitionSpeech Language ModelsAudio-Language ModelsLLM Self-Correction
AskPDF
Aug 31, 2026

Stride-k Subsampling: Train-Free Audio Token Reduction for Whisper

Whisper exposes speech through a fixed 1500-token encoder interface, now a default representation for ASR decoders and Whisper-based speech language models (SpeechLMs), yet its redundancy remains largely unexamined. We propose stride-k subsampling, a deterministic indexing operation that retains every k-th token after the convolutional stem or encoder transformer. Across five Whisper scales, k=2 preserves baseline WER at both positions, with CKA attributing this stability to acoustic overlap at the stem and attention-induced redistribution at the encoder output. Applying stride-2 at both positions cuts audio tokens by 75% and total GFLOPs by 52-58%, with small WER costs on most ASR benchmarks and larger costs on harder ones. The same configuration extends to three Whisper-based SpeechLMs, yielding modest accuracy drops on stronger baselines and larger drops on weaker ones, while reducing end-to-end latency by 19.6-27.4%. Requiring no training or auxiliary computation, stride-k subsampling exploits Whisper's preprocessing redundancy, indicating that its audio-token interface carries more capacity than downstream tasks require.

Chanhee Cho, Junhyuk Choi, Bugeun Kim

Department of Artificial Intelligence Chung-Ang University, Seoul, Republic of Korea

Audio Token CompressionAutomatic Speech RecognitionSpeech Language Models
AskPDF
Aug 29, 2026

HEAR Who Said What: Unlocking Speaker-Attributed Reasoning via Counterfactual Voice Grounding

Speech Language Models (SLMs) are increasingly deployed in multi-speaker environments, yet their ability to attribute speech to the correct speaker and reason over speaker identities remains unclear. Hence, we introduce HEAR, a conceptually hierarchical benchmark diagnosing the foundational capabilities of speaker-attributed reasoning, comprising 2.4K human-verified samples from 887 diverse multi-party audio clips. Evaluating 20 leading SLMs on HEAR reveals they struggle with these foundational tasks, often relying on semantic priors rather than actual vocal cues. To address this, we present A2R, a 30B model optimized on Counterfactual Audio with Speaker-level Hard negatives (CASH), a dataset designed to guide the model to prioritize acoustic vocal cues over linguistic signals. A2R achieves strong performance on HEAR and exhibits zero-shot generalization to diverse multi-speaker downstream tasks, demonstrating that learned speaker attribution unlocks the model's latent capacity for speaker-aware reasoning. All resources are available at https://attributetoreason.github.io/AttributeToReason/

Dongwook Lee, Sangkwon Park, Eunwoo Song +6

IPAI, Seoul National University (SNU) · Department of ECE, SNU · Department of EE, Yonsei University +3

Speech Language ModelsAudio-Language Model Evaluation
AskPDF
Aug 27, 2026

SURE-Voice: A Front-End Baseline for Speech-Evidence Filtering in Speech LLMs

Speech language models (speech LLMs) can generate plausible outputs from audio that contains no usable speech evidence. We study this failure as a pre-generation support-estimation problem and present SURE-Voice, a training-free front end that decides whether an audio prompt contains intelligible speech evidence before calling a speech LLM. We build SURE-Challenge with a 640-example SURE-Core split and a 1,920-example SURE-Extended split derived from 120 LibriSpeech source utterances. Using one fixed operating point, an energy screen plus Whisper token confidence raises unsupported accuracy on the held-out Extended test from 0.000--0.133 to 0.919 for six non-degenerate speech LLM backbones, while supported accuracy remains 0.919--0.970 and downstream calls fall from 480 to 287. A 500-clip ESC-50 sanity set shows the same pattern on real environmental audio, with vocal non-speech as a residual failure mode. An overlap diagnostic shows that source attribution remains separate from speech-evidence filtering. The evidence supports a controlled benchmark baseline and a deployment-oriented analysis; it does not establish universal robustness to semantic answerability, gain variation or natural conversations.

Mengzhe Geng

National Research Council Canada

Audio-Language Model EvaluationHallucination DetectionSpeech Language Models
AskPDF
Aug 26, 2026

VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction

Conversational systems, such as duplex speech language models (SLMs), still lack a streaming, accurate, and empathetic memory system as their soul. We introduce VoiceMem, a simple memory architecture with a parallel informational left brain, an emotional right brain, and streaming memory I/O mechanisms. We further build a complete pipeline for memory-aware SLM training, long-horizon evaluation, and decoupled deployment with interchangeable memory backends. Experiments and real-world deployment show three advantages: i) Accuracy: under top-5 retrieval, the left brain outperforms classical systems such as Mem0 at top-200 by nearly 30 points; ii) Emotional & Personal: the right brain, with short- and long-horizon affective attribution and dual-node persona modeling, achieves state-of-the-art performance across three persona benchmarks and improves the aggregate score by 4.29 points over the previous best system; and iii) Real-Time & Cheap: VoiceMem completes retrieval in 134 ms, well within standard VAD latency, adding no extra conversational delay while maintaining high accuracy and low cost. These results show that VoiceMem provides a practical memory foundation for real-time, personalized, and emotionally aware speech interaction.

Zhifei Xie, Jiaqi Lang, Ze An +7

Nanyang Technological University · Open Interaction Lab · National University of Singapore +2

Conversational MemoryLong-Term Conversational MemorySpeech Language ModelsFull-Duplex Spoken Dialogue Systems
AskPDF
Aug 25, 2026

Relative Time Intervals Representation for Word-level Timestamping with Masked Training

Although Speech Large Language Models (SpeechLLMs) excel at speech understanding and generation, their capacity for fine-grained, temporally aligned outputs remains underexplored. Our work addresses this gap by enabling SpeechLLMs to jointly model speech content and temporal structure, effectively transforming them from content understanding machines" into temporal-aware content understanding machines". Specifically, we replace traditional absolute timestamps with relative timestamps, achieving a more compact vocabulary and stronger generalization capabilities. To efficiently infuse timestamp prediction ability into pre-trained large language models, we introduce a hybrid fine-tuning strategy: full-parameter fine-tuning of the timestamp-augmented embedding layer and language model head, combined with LoRA fine-tuning of the decoder layers. Moreover, we design a masked timestamp training objective, preventing the model from over-relying on ground-truth timestamps, and thereby enhancing robustness against noisy real-world annotations. Extensive experiments demonstrate that our approach achieves significant improvements in timestamp prediction accuracy while maintaining strong speech transcription performance.

Quanwei Tang, Zhiyu Tang, Xu Li +3

Soochow University · University of Queenland · AISpeech Ltd +1

Speech Language ModelsAutomatic Speech Recognition
AskPDF
Aug 24, 2026

Do Spoken Language Models Hear Speech as They Read Text? Bridging Structural Gaps Between Speech and Text

Spoken Language Models (SLMs) generate textual responses directly from speech, offering an alternative to cascaded systems. Despite recent advances, existing SLMs still exhibit weaker instruction-following behavior and limited generalization across diverse tasks compared to text-based language models. Our analysis shows that speech and text representations in current SLMs remain weakly aligned despite strong downstream performance, indicating that structural differences between continuous, temporally varying speech and discrete text remain insufficiently addressed. To address this, we propose a simple framework that decouples length mismatch from semantic alignment and encourages closer correspondence between speech and text representations. Experiments across multiple benchmarks demonstrate competitive performance against strong baselines, underscoring the importance of explicitly addressing structural differences between speech and text in SLM training. Our code is publicly available at https://github.com/jaykim9870/Do_SLMs_Hear_Speech_as_They_Read_Text.

Hyeonyu Kim, Hwayeon Kim, Youngwon Choi +2

Maum AI Inc. · KAIST · Atmanity Inc.

Speech Language ModelsSpoken Language Understanding
AskPDF
Aug 24, 2026

DiaScriber: A Speech LLM for Joint Diarization and Transcription in Multi-Speaker Scenarios

Multi-speaker automatic speech recognition (MSASR) aims to jointly predict content transcriptions, speaker identities, and timestamps, thereby addressing the key question of "who spoke what and when" and holds substantial practical value in real-world multi-speaker scenarios. However, MSASR still encounters considerable challenges in the presence of fast turn transitions, overlapping speech, and complex, diverse multi-speaker scenarios. In this work, we propose DiaScriber, an end-to-end multi-speaker diarization and transcription model built on a speech large language model. We first construct diverse data pipelines to cover a wide variety of multi-speaker scenarios and their complexities, including validation and refinement, turn-transition and overlapping-speech simulation, and multimodal annotation. Furthermore, DiaScriber is developed based on the pretrained version of Qwen3.5-Omni through a three-stage training strategy involving continual pretraining, supervised fine-tuning, and reinforcement learning. Experiments show that DiaScriber achieves superior performance over comparison methods across extensive multi-speaker scenario test sets and demonstrates outstanding generalization ability in unseen multi-speaker scenarios.

Bingshen Mu, Xian Shi, Xiong Wang +6

Independent Researcher

Multi-Speaker ASRAutomatic Speech RecognitionSpeaker DiarizationSpeech Language Models
AskPDF
Aug 24, 2026

WnW: Waxing-and-Waning KV Cache for Long-Form Speech LLMs

Long-form audio inputs make the KV cache the dominant memory cost of speech LLMs. Prefill-only KV compression methods permanently discard audio KV positions once evicted, with no pathway to recover them during decoding. We show this is fragile on long-form audio: prefill attention concentrates near the audio start (an attention-sink effect), while decode-time attention distributes broadly, and the two rankings overlap weakly. We propose WnW (Waxing-and-Waning KV cache), which classifies KV-heads into anchor, tidal, and fixed roles via offline calibration. Anchor heads keep all audio KV on GPU and yield a decode-time signal of which audio region each token is read from; tidal heads keep a CPU-resident complement that is recalled chunk-by-chunk based on aggregated anchor-head scores; fixed heads keep only an on-GPU subset, with the rest permanently discarded. On LibriSpeech-Long with two 3B backbones (Voxtral-mini-3b and Qwen2.5-Omni-3B), WnW preserves near-Full-Cache accuracy while keeping only 20% of audio tokens on GPU, where prefill-only baselines fail to terminate. Results generalize across language, task, and domain shifts, and CPU-GPU recall adds little decode-time overhead in our measurements.

Yiming Yao, Chenyang Lyu, Xuanfan Ni +4

School of Informatics, Xiamen University · Alibaba Group

KV-Cache OffloadingKV-Cache CompressionLLM Inference AccelerationSpeech Language ModelsLong-Context Language Modeling
AskPDF
Aug 21, 2026

Grounded Decoding for Autoregressive Speech Enhancement via Adaptive Code-Space Grounding and Local LLM Refinement

Large language model (LLM)-based autoregressive speech enhancement (SE) produces natural speech using learned clean-speech priors, but may hallucinate content unsupported by the input. Deterministic SE better preserves observation-coupled evidence, yet often retains residual noise or local distortion. We propose an evidence-grounded generative SE framework that uses a deterministic estimate as imperfect evidence. A Whisper-guided DPRNN produces an enhanced waveform, which is blended with the observation and tokenized into a discrete evidence sequence. The evidence conditions an autoregressive clean-speech token generator and is reused during decoding through Code-Space Grounding (CSG), which penalizes candidates according to their Hamming distance in the factorized finite-scalar-quantized (FSQ) space. Because the appropriate grounding strength depends on acoustic difficulty, we introduce SNR-Conditioned CSG (SNR-CSG), which maps a calibrated residual-SNR estimate to an utterance-level strength and constructs an adaptive grounded anchor. Although grounding improves content fidelity, the anchor may retain local acoustic defects inherited from the evidence. Since such defects are predominantly local in the FSQ space, nearby tokens may provide better acoustic realizations without large departures from the observation-supported trajectory. We therefore propose Grounded Neighborhood Refinement with LLM ranking (GNR-LLM). It performs one additional teacher-forced pass conditioned on the grounded-anchor history, intersects the LLM top-KK candidates with a local FSQ Hamming neighborhood. Experiments on in-domain, controlled-SNR, and DNS no-reverb conditions show that SNR-CSG provides robust automatic grounding, while GNR-LLM substantially improves low-SNR perceptual quality without sacrificing content fidelity.

Hao Shi, Yuan Gao, Zhaoheng Ni +4

Kyoto University · Meta · Brno University of Technology +3

Speech EnhancementSpeech Language ModelsAutoregressive Decoding
AskPDF
Aug 21, 2026

Do SpeechLMs Hear Their Own Opinions? Diagnosing and Mitigating Previous-Belief Contamination in Streaming Emotion Understanding

Streaming emotion understanding uses historical state while continuously interpreting current audio, often feeding the model's previous prediction back as context. We show that this history conditioning can distort current perception. On a balanced CREMA-D-Stream counterfactual diagnostic, changing only the injected previous emotion label while holding the audio fixed reduces current-audio accuracy from 72.50% to 30.42% and flips 65.69% of predictions. The effect is strongly label-asymmetric, with prior pull ranging from 4.76% to 98.20%, revealing a failure we call previous-belief contamination (PBC). To address PBC, we introduce EmoUpdate, a training-free framework that separates current-audio perception from historical state revision through three components: (1) a prior-blind acoustic firewall that prevents historical state from entering perception; (2) an evidence-shrunk causal belief filter that introduces history only after observation formation and retains label-asymmetric transition structure only when supported by observed evidence; and (3) a closed-form decontamination operator derived from the same counterfactual measurements for serving stacks where firewalling is unavailable. Across four SpeechLMs and two streaming emotion benchmarks, EmoUpdate achieves the best step accuracy and state-balanced accuracy in all eight model--benchmark settings, improving S-BAcc by up to 69.71 points and step accuracy by up to 38.41 points over the strongest controlled baselines.

Haoyue Liu, Zhichao Wang, Ye Chen +2

School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen 518172, China · XJTU-POLIMI Joint School, Xi’an Jiaotong University, Xi’an 710049, China · University of California, Berkeley +1

Emotion Recognition in ConversationsSpeech Language ModelsBelief Updating in LLMs
AskPDF
Aug 20, 2026

Towards Quantifying Benchmark Optimization in ASR Models

Public benchmarks are important measures of Automatic Speech Recognition (ASR) model capabilities. However, by nature of being public, there is risk of models being optimized for these benchmarks in ways that do not generalize well to real-world data. We present a methodology for quantifying benchmark optimization, focusing on cases where the audio underdetermines the reference transcript. We identify three families of behavioral probes that reveal models' capabilities of reproducing benchmark reference spans despite underdetermined audio: reference disagreement, masked-number recovery, and orthographic switching. We find that the highest-scoring open source models output verbatim reference transcript spans even when the relevant audio is contradictory, masked, or ambiguous. Using a variety of mechanistic probes, we show that models respond to narrow acoustic cues to override the faithful representation of the audio in favor of a benchmark-optimized policy. We show the benchmark-optimized behavior can be causally manipulated via low-rank linear steering or simply appending audio to the end of a segment in some cases. Overall, our results indicate that high-performing models exhibit benchmark-conditioned behaviors that can inflate benchmark performance without reflecting improved general-purpose transcription ability.

Theo Lebryk, David Ayllon, Alice Baird +3

Hume AI Research

ASR EvaluationSpeech Language Models
AskPDF
Aug 13, 2026

CASA: Content-Acoustic Speaking Assessment with Speech Encoder and Large Language Model

Research on automatic speaking assessment (ASA) has increasingly adopted multimodal speech large language models to assess learners' speaking performance. However, existing studies provide limited analysis of how acoustic and content information contribute to predictions and how stable the resulting performance is. We propose CASA, a simpler architecture combining Whisper-medium and Qwen3.5-2B that achieves state-of-the-art performance while providing a more interpretable separation between speech delivery and content. On the Speak & Improve Corpus 2025, CASA achieves a root mean square error (RMSE) of 0.358, improving on the previous best RMSE while using approximately half the estimated inference parameters. The general-purpose architecture is designed for adaptation to other ASA corpora without structural changes and relies on three handcrafted fluency features. Through ablations and repeated runs, we analyze the individual and complementary contributions of acoustic and content information, examine performance variability, and demonstrate the potential of large language model reasoning for training-free content validation.

Nhan Phan, Ilona Lähteenmäki, Anna von Zansen +4

Department of Information and Communications Engineering, Aalto University, Espoo, Finland · Department of Education, University of Helsinki, Helsinki, Finland · Programmable Autonomous Systems Division, Walton Institute, Waterford, Ireland

Automated EvaluationSpeech Language ModelsAutomated Spoken Language Assessment
AskPDF