Speech Encoder

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).

Automatic speech recognition (ASR) systems exhibit unequal error rates across speaker groups, motivating interventions on their internal representations. We ask whether speaker-linked attributes that are linearly readable from pretrained ASR encoders yield useful directions for reducing group word-error-rate (WER) gaps. Across Whisper-medium, HuBERT-large, and Wav2Vec2-large on Common Voice and the Speech Accent Archive, we probe every encoder layer for metadata-derived sex/gender, age, and native/accent labels; construct centroid and probe-derived directions; inject them at selected layers; and compare downstream probe trajectories with matched WER changes. Sex labels are highly decodable (best macro-F1 0.924--0.941), native/accent labels are also above chance (0.544--0.696), and age is weaker (0.354--0.397). Of 22 post-selected reruns, nine have 95% paired-bootstrap intervals entirely below zero, yet every absolute source-group WER reduction is below 0.7 percentage points. Conversely, a local target-class probe rate can rise from 8.09% to 99.87% while WER worsens. Linear readability is therefore neither evidence of causal use nor a reliable mitigation method. Our results motivate evaluating speech-bias interventions jointly at representation, propagation, and task levels.

Speech Large Language Models (Speech-LLMs), typically built from a pre-trained speech encoder, a modality projector, and an LLM fine-tuned with Low-Rank Adapters (LoRA), have shown strong Automatic Speech Recognition (ASR) performance on general-domain speech. However, adapting them to domain-shifted speech, such as child or dialectal speech, remains challenging under limited target-domain data. Given the dominant role of the LLM in Speech-LLMs, with cross-entropy loss applied only at the LLM output, the speech encoder may receive insufficient adaptation to new acoustic conditions. In this paper, we propose Encoder Awakening via Adapters (EAVA), a simple yet effective domain-adaptive fine-tuning method for Speech-LLM-based ASR. First, lightweight adapters are inserted into each encoder layer and trained exclusively, enabling target-domain acoustic knowledge to be incorporated into the encoder while preserving its pre-trained knowledge. Second, the full model is jointly fine-tuned on the target domain with LoRA applied to the LLM. Experiments on three domain-shifted ASR datasets, covering child and dialectal speech, show that EAVA consistently outperforms vanilla fine-tuning and other baselines, achieving new state-of-the-art performance.

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.

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.

Recent advances in zero-shot text-to-speech (TTS) have substantially improved speech quality and voice cloning fidelity. However, many zero-shot TTS systems still depend on audio prompt transcripts at inference time. This dependency limits cross-lingual voice cloning, since in-the-wild reference audio is often untranscribed. In this technical report, we present Confucius4-TTS, a multilingual zero-shot TTS system that supports 14 languages and performs both intra-lingual and cross-lingual reference cloning without requiring transcripts of audio prompts. Confucius4-TTS follows a two-stage architecture, consisting of text-to-semantic (T2S) and semantic-to-acoustic (S2A) modules. The LLM-based T2S module uses a learnable speaker encoder to extract timbre features from self-supervised speech representations, and the conditional flow-matching S2A module converts the predicted semantic tokens into mel-spectrograms. The same model also supports continuation cloning when a reference transcript is available. Confucius4-TTS is trained on large-scale multilingual speech data. It achieves high intelligibility and speaker similarity on public benchmarks. On the CV3-Eval cross-lingual benchmark, Confucius4-TTS obtains an average WER of 3.73% across six directions. On our internal cross-lingual set, it achieves the best average overall rank in human evaluation among recent open-source and commercial systems. We release code, model checkpoints, and demos at https://github.com/netease-youdao/Confucius4-TTS.

Recent advances in model design and self-supervised audio representations have improved speech and audio understanding, yet infant-centered naturalistic recordings remain challenging due to limited labeled data, low signal-to-noise ratio, and cross-family domain shifts. We present a family-conditioned, multi-tier audio tagger that combines a LoRA-finetuned Whisper encoder with a lightweight, target-speaker-aware Transformer for long-context inference and framewise prediction across tiers. To improve temporal coherence, we incorporate a simple sequence-level smoothing loss, and to enhance robustness across households, we introduce a factorized speaker-token design with a shared tier token and a learned family-specific offset, reducing family bias and promoting generalizable representations. Together, these choices enable efficient and effective infant-centered audio tagging of daylong audio recordings in home environments.

A common approach to adding audio to a vision-language model is to train or adapt a large omni-modal system. We show that a lightweight alternative can be highly effective for music audio-visual question answering (AVQA). Qwen-MusicAVQA-7B connects a frozen Whisper encoder to Qwen2-VL-7B-Instruct through learned linear projections. The same frozen encoder processes both the video's music track and a TTS-spoken question through separate projectors, while the language model fuses visual frames, music, and question audio through pretrained self-attention, with no task-specific fusion network. On MUSIC-AVQA, our system reaches 96.0% +/- 3.9% accuracy across three independent training seeds on the 7,402-question available-video test subset. Our central finding is that downstream accuracy tracks how much fine-grained local temporal information the audio representation preserves. In a matched 32-token comparison, a stride-pooled Whisper frame sequence outperforms a globally pooled PANNs representation expanded to the same budget by 26 percentage points, even though PANNs sees at least as much audio and uses a far larger projector. The effect is not simply sequence versus vector: within Whisper alone, reducing temporal resolution at a fixed token budget costs a comparable amount. Under matched data and inputs, fine-tuned Qwen2.5-Omni-7B reaches 80.9%, against 95.9% for our 30 s variant; because the systems differ in backbone and adaptation, this is a system-level comparison. Accuracy remains high on sampled head and tail splits of the rephrased MUSIC-AVQA-R benchmark (96.5% and 95.6%). Because both encoders stay frozen and the music features are cached, the entire adaptation is cheap to train: the complete two-stage AVQA run takes approximately 5 hours on a single A100 80GB, and every run reported here fits on that one GPU.

Speech recordings often contain missing, corrupted, or incorrect regions that must be reconstructed or modified without re-synthesizing the entire utterance. Speech inpainting restores missing segments, whereas speech editing replaces spoken content according to an edited transcript. Both tasks require the generated speech to express the intended words while remaining consistent with the surrounding speaker identity, prosody, timing, and recording conditions. Discrete diffusion is particularly well suited to these tasks because it can iteratively refine masked tokens while jointly conditioning on both left and right acoustic context. We introduce SIEDD, a discrete diffusion framework for text-guided speech inpainting and editing over hierarchical codec tokens. Its core architecture, HiCoDD, follows the RVQ generation order by representing previously generated codebooks as clean, committed acoustic context and applying diffusion only to the current refinement codebook. This separation enables leakage-free joint training while matching sequential coarse-to-fine inference. The model further combines phoneme-level conditioning, span-localized classifier-free guidance, and duration prediction to support both fixed-duration inpainting and variable-duration text edits. On the RealEdit benchmark, SIEDD achieves the best overall speech-editing performance among the evaluated methods. It also outperforms the evaluated autoregressive baselines across all speech-inpainting settings, on both single and multiple gaps. These results demonstrate that explicitly modeling the codec hierarchy substantially improves context-preserving speech reconstruction and editing. See our full code at https://github.com/iftachShoham/SIEDD.

Multimodal large language models (MLLMs) have achieved significant success in speech-to-text translation (S2TT). However, when processing multilingual speech inputs, a single speech encoder shared across all languages suffers from the curse of multilinguality: languages at different resource levels compete for limited representation capacity, leading to strong high-resource performance but substantial degradation on low-resource speech. To address this problem and improve multilingual consistency, we propose MSRT, a novel framework built around a resource-aware Mixture of Speech Encoders (MoSE). MoSE uses an explicit language router to assign each utterance to an appropriate expert encoder. A frozen expert preserves high-resource language capabilities, while a trainable expert adapts to and specializes in medium- and low-resource languages. We further introduce a five-stage curriculum learning strategy that substantially reduces data dependence, requiring only 10 hours of paired S2TT data per language for effective alignment. We conduct extensive experiments on 45 languages, systematically evaluating all $45 \times 44$ translation directions. Our 4B-parameter model achieves state-of-the-art performance, outperforming substantially larger baselines. Empirical analyses show that MoSE improves high-, medium-, and low-resource languages simultaneously, with the largest gains on low-resource speech, thereby breaking the curse of multilinguality without compromising high-resource performance. To support future multilingual S2TT research, we release our code and models.

We present MERaLiON-GR, a speech gender recognition system that performs binary classification (female / male) on English and Southeast Asian (SEA) languages. The model finetunes MERaLiON-SpeechEncoder-2, a large conformer based transformer pre-trained on a broad speech corpus, and applies parameter efficient fine-tuning via Low-Rank Adaptation (LoRA) to adapt the encoder to the gender recognition task, and appends a multi-scale ECAPA-TDNN down stream network with attention pooling and a lightweight linear classifier. Extensive evaluations across multilingual Singaporean and Southeast Asian languages (English, Chinese, Malay, Tamil, Thai, Vietnamese, Indonesian, and Khmer) show that MERaLiON-GR consistently surpasses the state-of-the-art gender recognition model Vox-Profile and a large Audio-LLM, in both full-utterance and segment level evaluation modes. The results underscore the value of dedicated speech models in achieving accurate paralinguistic understanding and strong cross-lingual generalization.

Automatic speech recognition (ASR) performance degrades sharply on dysarthric speech, yet how disordered articulation reshapes a model's internal representations is underexplored. We present a layer-wise probing analysis of a transformer ASR encoder on Mandarin dysarthric speech under three transcript-matched conditions: original dysarthric speech, speaker conditioned zero-shot TTS resynthesis, and unconditioned TTS. The probes reveal a task-dependent hierarchy: phoneme boundary information stays weak for dysarthric speech at every layer, phoneme identity becomes recoverable toward the upper layers, and recognition difficulty is encoded in the deepest layers. Tone-sensitive evaluation shows Mandarin lexical tone is a persistent error source. Cross-condition similarity divergence grows with depth, indicating that disordered speech affects high-level representations more than low-level acoustic features. Guided by these findings, single-layer LoRA at layer 7 and adaptation on subset layers 5-8 achieve performance within 3.5% and 2.48% relative margins of full encoder adaptation, respectively, while upper-layer adaptation is less effective for dysarthric speech. These findings link representation analysis to parameter-efficient fine-tuning and motivate layer-aware adaptation for low-resource Mandarin dysarthric ASR.

Speech deepfake detection has expanded in scope with increasingly heterogeneous spoofing mechanisms, including speech synthesis, voice conversion, vocoder reconstruction, and neural-codec resynthesis. The resulting spoofing artifacts can be further shaped by variability in source speech, recording environments, and transmission channels. This variability makes robust generalization across heterogeneous conditions a central requirement for practical detection systems. This report presents Teffic-Audio, a general speech deepfake detection system designed for comprehensive evaluation environment. Teffic-Audio adopts a straightforward detector architecture consisting of a Conformer-based speech encoder, multi-head attentive statistics pooling, and a binary classifier. Rather than relying on additional architectural complexity, the system improves generalization through its training recipe, which integrates multi-source data, attack- and source-balanced sampling, and diverse audio augmentation. Trained only with open-source data, Teffic-Audio achieves a pooled EER of 1.454% on the 14 test sets of Speech-DF-Arena, outperforming all currently public systems on the leaderboard. It also obtains the lowest EER on five individual test sets and shows a favorable performance-complexity trade-off compared with larger leading systems. Overall, Teffic-Audio provides a strong and practical reference system for general speech deepfake detection.

Federated learning (FL) enables privacy-preserving training of automatic speech recognition (ASR) systems across distributed data sources, yet its application to large-scale speech language models (SpeechLLMs) remains unexplored. This paper presents the first systematic study of federated training for SpeechLLM-based end-to-end ASR systems. We design a communication-efficient federated optimization strategy tailored to the unique challenges of SpeechLLM architectures, addressing high-dimensional parameter spaces, gradient communication overhead, and computational constraints in distributed settings. Through extensive empirical evaluation on monolingual ASR tasks in English and Italian, we demonstrate the effectiveness and stability of our federated approach compared to centralized training baselines across diverse acoustic conditions and speaking styles. Additionally, we conduct a comprehensive ablation study analyzing the impact of different speech encoder architectures on monolingual English ASR performance within the federated framework, providing insights into optimal model configurations for decentralized training. Our results achieve competitive word error rates while reducing communication costs, establishing practical foundations for federated SpeechLLM deployment in real-world multilingual scenarios.

Lightweight projectors are an established way to connect pre-trained speech encoders with large language models (LLMs), mapping acoustic features into token-level embeddings for tasks like ASR and spoken question answering. Existing systems, however, typically only support a few languages and are often limited to English. We introduce MEUSLI, the first open-science multilingual projector family that links a Whisper encoder with open-source multilingual LLMs, enabling fully open-source end-to-end ASR in 28 European languages. MEUSLI extends prior monolingual pipelines, delivering strong results across high- and low-resource languages. Using proper continual leaning techniques, MEUSLI can be easily extended to other languages not seen in training. We further demonstrate that the MEUSLI projector can be leveraged beyond ASR, enabling multilingual speech translation and topic identification with only a few hours of task specific supervision per language. Overall, MEUSLI provides a solid foundation for multilingual speech understanding tasks, supporting scalable and inclu- sive open-source SpeechLLM

End-to-end speech language models increasingly represent user speech with speech tokens rather than relying exclusively on cascaded ASR--LLM--TTS pipelines. Although these tokens support expressive and low-latency spoken interaction, they may also preserve sensitive speaker characteristics. We investigate whether exposed speech tokens leak voiceprints and formulate this risk as a speaker inversion attack. We introduce Audio BERT (AuB), a trainable model that constructs token embeddings from discrete codebooks and aggregates them into speaker-sensitive representations, and propose SpInv, a two-stage inversion method built on AuB to recover embeddings in the space of an attacker-specified speaker encoder. We evaluate Moshi, Higgs3, Kimi-Audio, and Qwen3-Omni using speaker-disjoint protocols on the VoxCeleb dataset. Extensive experiments show that, with only three seconds of frontend output, SpInv achieves cosine similarities above 0.70 in the attacker-specified speaker-encoder space.

A voice actor's voice is their asset, and AI cloning directly threatens it. The natural defense flags the enrolled actor whose embedding similarity to a suspect recording crosses a threshold. We show it fails where it is most needed: trained voices crowd the embedding space, and each actor performs many styles. On 1,168 Japanese voice actors (56,568 segments, ~63 h), a misidentification floor survives calibration, score normalization, and discriminative re-ranking (linear and nonlinear, including PLDA): the residual is a limit of the embedding geometry, not of the back-ends we evaluate. The best ensemble still leaves ~2.6% closed-set misidentification, several-fold above matched controls; session-disjoint, re-ranking lowers the floor only to 13.0%. The same crowding drives false attribution: on a generic English encoder, roughly half the clones of non-enrolled people falsely accuse an enrolled actor, while -- by a separate real-vs-synthetic shift -- 32% of Seed-VC clones of enrolled targets are missed at the same threshold; one operating point couples the two, and none escapes both. A domain-matched, voice-actor-trained encoder mitigates substantially (a four-fold gender gap vanishes; wrongful misattribution falls to 1.5-10%), but does not remove the floor. Controls (codec, channel, vocoder, content) support reading the miss rate as a real-versus-synthetic covariate shift, not missing speaker information. Fixed-threshold clone attribution is thus unreliable here, and on a generic encoder unfair. Robust attribution must extend spoofing-aware speaker verification to open-set 1:N (anti-spoofing gate, domain-matched encoder, per-speaker calibration, abstain option), and even then supports detection, not autonomous enforcement.

Automatic speech recognition is dominated by autoregressive decoders that emit one token at a time. We ask whether a discrete diffusion language model can transcribe speech instead, refining a whole transcript in parallel over a small number of denoising steps. We train an audio-native interface for DiffusionGemma, a 26B mixture-of-experts model that generates text by uniform, random-token discrete diffusion rather than the absorbing-mask scheme common to recent diffusion language models. A frozen Whisper encoder supplies acoustic features, a lightweight projector maps them into the model embedding space, and low-rank adapters let the frozen backbone attend to the new modality. About 42M parameters are trained, which is 0.16 percent of the backbone. We find that the natural training objectives fail to ground the audio because their gradient reaches the projector only through attention that has already dismissed it. A connectionist temporal classification loss applied through the frozen output head breaks this deadlock. The resulting model reaches 6.6 percent word error rate on LibriSpeech test-clean, transcribes in roughly eight parallel steps regardless of utterance length, and uses a single adapter trained on six languages, which we evaluate here on English, Hindi, and Mandarin.

Single-channel speech separation remains challenging for real-world deployment due to source permutation ambiguity, sampling variability of generative models, and the difficulty of processing long recordings with chunk-wise inference. We address these issues with a conditional flow-matching-based method that produces an ordered two-source output conditioned on the mixture. A frozen speaker encoder defines the source order during training and is reused at inference for biometric best-of-$N$ candidate selection and chunk-level channel alignment. We evaluate separation quality on Libri2Mix benchmark using SI-SDR, PESQ, and ESTOI, and measure downstream impact using cpWER for automatic speech recognition and EER for speaker verification. The results show that the proposed Transformer U-Net variant is competitive with strong baselines in objective separation metrics and achieves the lowest downstream automatic speech recognition and speaker verification error rates in all evaluated settings.

Audio-language models compress a speech encoder's output through a Querying Transformer (Q-Former) connector before feeding it to a large language model. We identify two failures in this compression. The connector's output vectors collapse to a single direction, and different speakers produce nearly indistinguishable outputs, with paralinguistic cues such as speaker identity, gender, and prosody lost along the way. Our method, ORCA, reverses this collapse by splitting the queries into groups whose outputs are constrained to point in different directions. On SAKURA multi-hop reasoning, ORCA gains 26.4 points over an identically trained 4B baseline, reaching 75.2% (vs. 49.0% for the 8B Audio Flamingo-3). At the connector level, the same change cuts query redundancy by 12x and raises cross-speaker variance by 75x.

We re-implement the NAVER LABS IWSLT 2025 instruction-following pipeline for the IWSLT 2026 Shared Task (constrained condition, short audio track), adapting it to the mandated components: SeamlessM4T-v2-large as the speech encoder and Qwen3-4B-Instruct as the LLM backbone. The three-stage approach projector alignment, text-only LoRA pre-training, and multimodal merging is preserved from the original design. We additionally construct 100k synthetic instruction-following examples across ten speech-centric task types (10k per task) from the provided corpora, suitable for further Stage 3 fine-tuning. Our primary model achieves COMET 0.781 on EN-ZH speech translation and BERTScore-F1 0.346 on English SQA on the MCIF benchmark.

Multi-talker Automatic Speech Recognition (MTASR) faces significant challenges in accurately transcribing overlapping speech, particularly under complex high-overlap conditions. While recent Mixture-of-Experts (MoE) approaches have shown promise, they typically rely on frame-independent routing that leads to temporal myopia, and depend solely on the downstream ASR objective, which results in implicit and ungrounded representation learning. To address these limitations, we propose Holistic Speaker-Aware Guided Experts (H-SAGE) for MoE-based MTASR. Specifically, we introduce a Speaker-Aware Global Encoder to capture long-term dependencies, supervised by an auxiliary Overlap-Aware Loss that explicitly guides the model to discern acoustic states. Furthermore, we design a Holistic Gating Mechanism to arbitrate expert selection by jointly evaluating global context and local details. Experiments on LibriSpeechMix demonstrate that H-SAGE achieves consistent improvements over strong baselines, particularly in complex scenarios, validating that explicit acoustic guidance effectively enhances expert collaboration. Our code can be found at https://github.com/NKU-HLT/H-SAGE.

Fueled by increasing model scale and multimodal inputs, Multimodal Large Language Models (MLLMs) have emerged as a promising paradigm for Spoken Language Assessment (SLA). While effective, this paradigm often overlooks the intrinsic ordinal structure of language acquisition. This paper works around the necessity of large-scale MLLMs by introducing Latent Ordinal Prototype Alignment (LOPA) for SLA, a prototype-based regularizer that enforces an ordinal geometric prior directly on the latent space. Coupled with Semantic-Anchored Layer Routing (SALR), which adaptively harvests multi-depth representations from a frozen Whisper encoder, our framework achieves an RMSE of 0.361. This performance rivals billion-parameter systems without the need for LLM-based fine-tuning. Further analysis reveals that SALR's synergy with LOPA offers interpretable, criterion-aligned preferences, thereby supporting an efficient and ordinal-aware modeling alternative to current scaling-centric models for SLA.

Connecting a pre-trained speech encoder to a Large Language Model (LLM) is the standard architecture for building Speech LLMs. However, a structural misalignment exists between the encoder and the LLM. Unlike encoders based on automatic speech recognition, which often produce representations in separate language-specific spaces, LLMs operate within a unified language-agnostic space. A mechanism is required to align the encoder's language-specific representations with the LLM's shared space. We argue that speech translation provides a principled way to achieve this. Unlike monolingual transcription, translation requires the model to bridge different languages and learn language-agnostic representations. We experimentally evaluate the impact of incorporating translation objectives into speech encoder pre-training. Our results demonstrate that translation-enhanced pre-training improves cross-modal integration and leads to superior performance across downstream Speech LLM tasks.

Adapting a streaming speech recognition model to a new language requires choosing between two plausible warm starts: a multilingual (ML) encoder or an English-only (EN) encoder. The common intuition is that the multilingual encoder should help most at low data, but it is unclear how long that advantage persists, whether tight streaming latency amplifies it, and whether it survives deployment quantization. We answer these questions with a controlled sweep of a 0.6 B-parameter cache-aware FastConformer transducer across eight European languages, up to five target-language data scales (100 h to 2500 h), three streaming tiers plus offline decoding, and up to four public test sets. The main result is that multilingual initialization is a data-limited advantage, not a latency-limited one. On FLEURS at 160 ms, the mean EN-ML word error rate (WER) gap falls from +4.21 percentage points (pp) at 100 h to +0.20 pp at 2500 h; a power-law fit summarizes this decay, with each doubling of target-language data roughly halving the remaining advantage. Across the three streaming tiers, the across-language mean EN-ML gap is approximately stable at each scale from 100 to 1000 h, and is near zero by 2500 h. Finally, 4-bit weight-only encoder quantization at the matched 560 ms streaming tier reduces the encoder footprint by about 3x, with an average FLEURS WER increase of about 0.5 pp. The resulting guideline is simple: use multilingual initialization in low-data regimes, treat the choice as effectively irrelevant at large data, and make latency and quantization decisions independently.

The mismatch between continuous acoustic signals and discrete event-driven processing remains a fundamental bottleneck for neuromorphic speech processing. Current systems typically rely on fixed spike encoders, forcing downstream Spiking Neural Networks (SNNs) to compensate for non-adaptive input representations. To address this, we present a learnable residual speech-to-spike encoder jointly trained end-to-end with a Recurrent Leaky Integrate-and-Fire (R-LIF) backbone. We validate this approach on the Google Speech Commands v2 (GSC-v2) benchmark, achieving up to 94.97% accuracy. Notably, the learned encoder remains highly parameter-efficient with a compact 35k-parameter variant that reaches 89.8%, matching or exceeding prior baselines that require an order of magnitude more parameters. Our encoder-focused analysis, including linear probing and gradient-residual inspection, indicates that the encoder does not target faithful signal reconstruction but instead learns task-aligned spike representations that enhance class separability. Finally, we benchmark bio-inspired, hardware-friendly credit assignment by comparing Direct Feedback Alignment (DFA) with surrogate-gradient BPTT under identical architectures and training conditions. We find that DFA reaches 91.5% accuracy, quantifying the performance trade-off of bio-inspired learning rules for modern neuromorphic audio.

This paper introduces CraBERT, a pre-trained phoneme encoder (PPEnc) designed for efficient pre-training in text-to-speech (TTS). CraBERT employs a cascade-fusion architecture and a subword-phoneme alignment algorithm to integrate representations from a pre-trained subword-level BERT into a phoneme-level BERT. This design provides prior word- and sentence-level information, reducing the amount of pre-training required by the phoneme encoder. Subjective listening evaluations show that CraBERT achieves MOS values comparable to existing PPEncs after approximately one epoch of pre-training, whereas the baselines in our comparison are pre-trained for approximately ten epochs. These results demonstrate that CraBERT can efficiently learn representations suitable for improving the perceived naturalness and prosody of synthesized speech.

Zero-shot text-to-speech (TTS) relies on robust speech representations. However, current speech tokenizers face a fundamental trade-off: acoustic codecs preserve high-fidelity audio but lack linguistic constraints, causing content errors during generation, whereas semantic tokens from self-supervised learning (SSL) models ensure precise text alignment but discard some acoustic information. To bridge this gap, we propose SARA, a dual-stream VAE that directly fuses a frozen SSL semantic anchor with a dedicated residual acoustic encoder. This effectively mitigates the dilemma, creating an efficient and compact latent space without relying on complex regularizers. SARA achieves superior reconstruction quality over strong baselines. Furthermore, in downstream zero-shot TTS tasks, it yields highly natural and expressive synthesis quality, and maintains robust generation performance even under accelerated inference, offering a favorable trade-off between synthesis speed and computational cost.

Speech foundation models often struggle in low-resource domains due to domain mismatch and data scarcity. We propose Gumbel-BEARD, a domain adaptation framework that automates Whisper encoder layer selection via an end-to-end trainable hard Gumbel-Softmax selector. It enables self-supervised adaptation with a BEST-RQ objective that dynamically adapts to target acoustic characteristics without manual tuning. Experiments on the MyST child speech corpus demonstrate efficiency and scalability: with 10 h of labeled data for fine-tuning, our method matches a fully supervised baseline trained on the complete 133 h labeled set. We establish new state-of-the-art word error rates (WERs) of 8.21% using Whisper-medium on MyST and 11.06% using Whisper-small on the OGI Spontaneous dataset. Evaluation on CORAAL further confirms robustness to adult dialectal domain shifts, with up to 6% relative WER reduction, highlighting the generalizability of our approach to diverse low-resource conditions.

Recent efforts to extend large language models (LLMs) to speech inputs typically rely on cascaded ASR-LLM pipelines, end-to-end speech-language models, or bridge/distillation-based adaptation. While these routes respectively reuse strong pretrained components, enable native speech-language interaction, or offer lightweight adaptation, they often suffer from transcript-interface latency, costly multimodal training, or sequential speech-language coupling. To address these limitations, we present AuRA, a method that distills audio encoding capability into the LLM. Specifically, AuRA feeds the same speech input to an ASR encoder (as a teacher) and a LoRA-adapted LLM (as a student) through a lightweight audio embedding layer, and uses layer-wise distillation to align the student's hidden states with corresponding teacher representations, thereby internalizing speech representations into lightweight LLM-side adaptations. Compared with cascaded and serial bridge methods, AuRA enables tighter speech-language joint modeling and efficient parallel end-to-end inference, while also reusing pretrained speech and language models rather than requiring large-scale multimodal training. On multiple speech-language benchmarks, AuRA consistently outperforms cascaded systems, speech-to-LLM adaptation baselines, and large-scale speech-language and multimodal models in both effectiveness and efficiency.

Speech-to-text (S2T) systems for recognition (ASR) and translation (S2TT) typically generate discrete text tokens. In contrast, continuous-target language modelling performs generation in a continuous space, yet its potential for S2T remains unexplored. To bridge this gap, we propose ELF-S2T, an audio-conditioned continuous-target generative model for S2T. Built upon the pre-trained Embedded Language Flows (ELF) backbone, ELF-S2T processes speech via a frozen Whisper encoder and a single linear projector, prepending the resulting audio condition to the noisy text latent for in-context, flow-matching denoising. To prevent the model from over-relying on its pre-trained text context, we introduce audio forcing during training, and further amplify the audio condition via classifier-free guidance at inference. Experiments on LibriSpeech and CoVoST2 show that ELF-S2T achieves competitive ASR and S2TT performance. Crucially, our error analysis reveals that, although ASR and S2TT errors look very different on the surface, both stem from the same underlying cause, a close distance confusion in the continuous latent space. This finding naturally aligns with the continuous representation generation paradigm, indicating a common semantic mapping process beneath recognition and translation. Our code and pretrained models are publicly available at https://github.com/Sslnon/ELF-S2T.

Automatic speech recognition (ASR) has advanced remarkably for standard speech; however, pathological speech from neurological conditions remains a significant challenge. We investigate speaker conditioning via Feature-wise Linear Modulation (FiLM), injecting x-vector-derived information into each transformer layer of a frozen ASR encoder to adapt internal representations to individual pathological speakers without modifying base model weights. We benchmark this for the ASR task against standard and parameter-efficient fine-tuning baselines, complemented by post-processing, on Spanish and English pathological speech. Additionally, we evaluate if the adapted model preserves the ability to answer speech-related questions. Results show that speaker-conditioned ASR is competitive with established adaptation strategies while retaining performance on non-conditioned speech.

We present Echo, a proof-of-concept audio system built around a single 25 M-parameter ViT encoder. The encoder is pretrained with a JEPA objective and then specialised by stages to carry speaker identity, phonetic content, and dynamic source routing in the same 512-dimensional latent space, with no per-task fine-tuning at deployment. Light heads handle diarization (ArcFace + VBx) and dynamic source separation (null-target K-set prediction). On synthetic VoxCeleb2 mixtures with unknown K, the canonical stack reaches 15.00% blind DER, 97.80% PIT separation accuracy with +9.52 dB latent SI-SDR, and a +53.50-point speaker/content factorisation gap on a held-out k-NN probe. The point of Echo is not a new SOTA on any single task but the joint coexistence of three tasks on one encoder at this footprint. We document the design stage by stage, report the dead-ends, and identify the structural wall on end-to-end ASR through the VQ bottleneck that still bounds the PoC.

We describe the winning system for Task 2 of the KSAA-2026 Shared Task on Arabic Speech Dictation with Automatic Diacritization. The task requires producing fully diacritized Arabic text from speech audio and undiacritized transcripts, with only 2,327 training samples available and no external data permitted. Our system fine-tunes CATT-Whisper, a character-level multimodal model combining a pretrained CATT text encoder with a frozen Whisper speech encoder. The key to our approach is training regularization: R-Drop consistency regularization, Optuna-optimized hyperparameters with high weight decay, and Focal Loss. At inference, we average 200 stochastic forward passes across four model checkpoints using Monte Carlo Dropout at the softmax probability level. The system achieves 23.26% WER on the primary leaderboard metric (with case endings, including no-diacritic positions), placing 1st among all participants.

Non-intrusive intelligibility prediction estimates how well hearing-impaired listeners understand hearing-aid-processed speech without a clean reference. We study this task in the 3rd Clarity Prediction Challenge using two frozen speech encoders, Canary and WavLM. The central question is not only whether complementary pretrained representations should be combined, but where their interaction should occur. We compare single-backbone baselines, uniform score averaging, pool-late fusion, cross-attention, frame-aligned fusion, and reverse alignment under a shared left/right-preserving binaural framework. Among the compared systems, the best model temporally prepares WavLM with a learnable strided convolution and fuses it with Canary on the coarser Canary timeline before pooling, reaching Eval RMSE 24.96$\pm$0.06 and Eval Corr 0.796$\pm$0.001. Severity, enhancement-system, layer-window, and temporal-shift analyses indicate that coarse local temporal correspondence before pooling is a useful inductive bias for this task.

We address text-assisted speech intelligibility prediction for hearing-impaired listeners in CPC3. Although the target is a sentence-level percentage, it is determined by reference-word recognition outcomes. We formulate prediction as reference-conditioned word-level correctness modeling: a frozen Whisper encoder analyzes degraded speech, a teacher-forced decoder conditions on the canonical transcript, and sentence intelligibility is obtained by averaging predicted correctness probabilities over valid reference words. To complement transcript-conditioned decoder states, we add a word-aligned local acoustic branch based on character-level cross-attention alignment and an utterance-level global acoustic branch for calibration. On the official evaluation set, the decoder baseline obtains RMSE 24.92 and correlation 0.795, while joint fusion improves to incorrect-word F1 0.778, MCC 0.626, correlation 0.806, and RMSE 24.39. A similar trend with Whisper medium suggests that the gain comes from prediction granularity and alignment-aware fusion.

In bandwidth-constrained communication such as satellite and underwater channels, speech must often be transmitted at ultra-low bitrates where intelligibility is the primary objective. At such extreme compression levels, codecs trained with acoustic reconstruction losses tend to allocate bits to perceptual detail, leading to substantial degradation in word error rate (WER). This paper proposes ClariCodec, a neural speech codec operating at 300 bit per second (bps) that reformulates quantisation as a stochastic policy, enabling reinforcement learning (RL)-based optimisation of intelligibility. Specifically, the encoder is fine-tuned using WER-driven rewards while the acoustic reconstruction pipeline remains frozen. Even without RL, ClariCodec achieves 4.64% WER on the LibriSpeech test-clean set at 300 bps, already competitive with codecs operating at higher bitrates. Further RL fine-tuning reduces WER to 3.55% on test-clean and 10.4% on test-other, corresponding to a 23% relative reduction while preserving perceptual quality.

Automatic detection of speaker confidence is critical for adaptive computing but remains constrained by limited labelled data and the subjectivity of paralinguistic annotations. This paper proposes a semi-supervised hybrid framework that fuses deep semantic embeddings from the Whisper encoder with an interpretable acoustic feature vector composed of eGeMAPS descriptors and auxiliary probability estimates of vocal stress and disfluency. To mitigate reliance on scarce ground truth data, we introduce an Uncertainty-Aware Pseudo-Labelling strategy where a model generates labels for unlabelled data, retaining only high-quality samples for training. Experimental results demonstrate that the proposed approach achieves a Macro-F1 score of 0.751, outperforming self-supervised baselines, including WavLM, HuBERT, and Wav2Vec 2.0. The hybrid architecture also surpasses the unimodal Whisper baseline, yielding a 3% improvement in the minority class, confirming that explicit prosodic and auxiliary features provide necessary corrective signals which are otherwise lost in deep semantic representations. Ablation studies further show that a curated set of high confidence pseudo-labels outperforms indiscriminate large scale augmentation, confirming that data quality outweighs quantity for perceived confidence detection.

Understanding the internal machinations of deep Transformer-based NLP models is more crucial than ever as these models see widespread use in various domains that affect the public at large, such as industry, academia, finance, health. While these models have advanced rapidly, their internal mechanisms remain largely a mystery. Techniques such as Sparse Autoencoders (SAE) have emerged to understand these mechanisms by projecting dense representations into a sparse vector. While existing research has demonstrated the viability of the SAE in interpreting text-based Large Language Models (LLMs), there are no equivalent studies that demonstrate the application of a SAE to audio processing models like Automatic Speech Recognizers (ASRs). In this work, a SAE is applied to Whisper, a Transformer-based ASR, training a high-dimensional sparse latent space on frame-level embeddings extracted from the Whisper encoder. Our work uncovers diverse monosemantic features across linguistic and non-linguistic boundaries, and demonstrates cross-lingual feature steering. This work establishes the viability of a SAE model and demonstrates that Whisper encodes a rich amount of linguistic information.

Integrating speech understanding and generation is a pivotal step toward building unified speech models. However, the different representations required for these two tasks currently pose significant compatibility challenges. Typically, semantics-oriented features are learned from self-supervised learning (SSL), and acoustic-oriented features from reconstruction. Such fragmented representations hinder the realization of truly unified speech systems. We present WavCube, a compact continuous latent derived from an SSL speech encoder that simultaneously supports speech understanding, reconstruction, and generation. WavCube employs a two-stage training scheme. Stage 1 trains a semantic bottleneck to filter off-manifold redundancy that makes raw SSL features intractable for diffusion. Stage 2 injects fine-grained acoustic details via end-to-end reconstruction, while a semantic anchoring loss ensures the representation remains grounded within its original semantic manifold. Comprehensive experiments show that WavCube closely approaches WavLM performance on SUPERB despite an 8x dimensional compression, attains reconstruction quality on par with existing acoustic representations, delivers state-of-the-art zero-shot TTS performance with markedly faster training convergence, and excels in speech enhancement, separation, and voice conversion tasks on the SUPERB-SG benchmark. Systematic ablations reveal that WavCube's two-stage recipe resolves two intrinsic flaws of SSL features for generative modeling, paving the way for future unified speech systems. Codes and checkpoints are available at https://github.com/yanghaha0908/WavCube.

Speech large language models (SLMs) are typically built from text large language model (TLM) checkpoints, yet they still suffer from a substantial modality gap. Prior work has mainly attempted to reduce this gap from the output side by making speech generation more text-like, but the gap remains. We argue that the key remaining bottleneck lies on the input side. We propose TextPro-SLM, an SLM that makes spoken input more closely resemble that of a prosody-aware text LLM. TextPro-SLM combines WhisperPro, a unified speech encoder that produces synchronized text tokens and prosody embeddings, with an LLM backbone trained to preserve the semantic capabilities of the original TLM while learning paralinguistic understanding. Experiments show that TextPro-SLM achieves the lowest modality gap among leading SLMs at both 3B and 7B scales, while also delivering strong overall performance on paralinguistic understanding tasks. These gains are achieved with only roughly 1,000 hours of LLM training audio, suggesting that reducing the modality gap from the input side is both effective and data-efficient.

A speaker encoder used in multilingual voice cloning should treat the same speaker identically regardless of which script the audio was uttered in. Off-the-shelf encoders do not, and the failure is accent-conditional. On a 1043-pair Western-accented voice corpus across English, Hindi, Telugu, and Tamil, WavLM-base-plus-sv loses 0.082 absolute cosine similarity when the same voice changes script and ECAPA-TDNN loses 0.105. On a 1369-pair Indian-accented voice corpus, the gap shrinks to 0.006 (WavLM-SV) and 0.044 (ECAPA-TDNN). The leak is largest where it matters most for cross-script TTS: when a system projects a non-Indic-trained voice into Indic scripts. We present LASE (Language-Adversarial Speaker Encoder), a small projection head over frozen WavLM-base-plus trained with two losses: a supervised contrastive loss over voice identity, and a gradient-reversal cross-entropy against a 4-language classifier that pushes the embedding to be language-uninformative while remaining speaker-informative. Trained on 1118 quality-gated cross-script pairs synthesised from 8 commercial multilingual voices, LASE's residual gap is consistent with zero on both corpora (Delta = 0.013 Western, Delta = 0.026 Indian; both bootstrap 95% CIs include zero) and amplifies the cross-script-vs-floor margin 2.4-2.7x over both baselines. An ECAPA+GRL ablation shows the GRL objective improves either backbone but the WavLM choice contributes too. In synthetic multi-speaker diarisation, LASE matches ECAPA-TDNN on cross-script speaker recall (0.788 vs 0.789) with ~100x less training data. We release the r1 checkpoint, both corpora, and the bootstrap recipe.

Modern automatic speech recognition (ASR) systems have been observed to function better for certain speaker groups (SGs) than others, despite recent gains in overall performance. One potential impediment to progress towards fairer ASR is a more nuanced understanding of the types of modeling errors that speech encoder models make, and in particular the difference between the structure of embeddings for high-performance and low-performance SGs. This paper proposes a framework typifying two types of error that can occur in modeling phonemes in ASR systems: random error/high variance in phoneme embedding, vs systematic error/embedding bias. We find that training phoneme classification probes only on a single, typically disadvantaged SG, sometimes improves performance for that SG, which is evidence for the existence of SG-level bias in phoneme embeddings. On the other hand, we find that speakers and SGs with higher levels of phoneme variance are the same as those with worse phoneme prediction accuracy. We conclude that both types of error are present in phoneme embeddings and both are candidate causes for SG-level unfairness in ASR, though random error is likely a greater hindrance to fairness than systematic error. Furthermore, we find that finetuning encoder models using a fairness-enhancing algorithm (domain enhancing and adversarial training) changes neither the benefits of in-domain phoneme classification probe training, nor measured levels of random embedding error.

In bandwidth-constrained communication such as satellite and underwater channels, speech must often be transmitted at ultra-low bitrates where intelligibility is the primary objective. At such extreme compression levels, codecs trained with acoustic reconstruction losses tend to allocate bits to perceptual detail, leading to substantial degradation in word error rate (WER). This paper proposes ClariCodec, a neural speech codec operating at 200 bit per second (bps) that reformulates quantisation as a stochastic policy, enabling reinforcement learning (RL)-based optimisation of intelligibility. Specifically, the encoder is fine-tuned using WER-driven rewards while the acoustic reconstruction pipeline remains frozen. Even without RL, ClariCodec achieves 3.68% WER on the LibriSpeech test-clean set at 200 bps, already competitive with codecs operating at higher bitrates. Further RL fine-tuning reduces WER to 3.20% on test-clean and 8.93% on test-other, corresponding to a 13% relative reduction while preserving perceptual quality.

Speech and audio encoders developed over years of community effort are routinely excluded from video understanding pipelines, not because they fail, but because benchmarks never required listening. We audit 10 video benchmarks and find items largely solvable from visual cues alone: a single-frame probe answers about 76% of AVQA without audio, suggesting poor measurement of audio-visual reasoning. Building on LLaVA-OneVision, we attach a speech/audio encoder and compare five compressor architectures under 25-fold token reduction (25 Hz to 1 Hz). Across 10 benchmarks, with and without filtering, audio yields clear gains on tasks requiring speech comprehension or cross-modal grounding, while vision-centric suites remain largely unaffected. Our results show that speech encoders play a larger role in video understanding than current benchmarks suggest. We will open-source our work at https://github.com/naver-ai/unimambamia-av.

Sparse Autoencoders (SAEs) decompose language-model activations into sparse, interpretable features, but standard encoders usually treat the latent dictionary as a flat set of independent coordinates, leaving hierarchy and feature interactions to emerge only implicitly. We propose KronSAE, a design that factorizes the latent space into heads and forms post-latent features as pairwise compositions of lower-dimensional pre-latents using mAND, a differentiable AND-like interaction. This imposes a compositional co-activation prior while remaining compatible with standard SAE objectives and variants such as TopK, Matryoshka, and Switch SAEs. KronSAE matches strong baselines on the EV-FLOPs frontier, improves interpretability of the latents, better captures the underlying correlated feature structure, and reduces encoder computational cost as an additional benefit. Code is available at https://github.com/corl-team/kronsae.