Dissecting Sensitivity to Training Language in Self-Supervised Speech Learning Using Neural Audio Codec Tokens

Neural audio codecs (NACs) have become popular for obtaining speech representations as discrete tokens. Beyond compression, discrete tokens can be used to train self-supervised learning (SSL) models. Such models, referred to as codec-based SSL models, reduce data storage and computational cost, enabling scalable SSL pre-training. However, their language sensitivity remains unclear. When the language changes, codec-based SSL models may require retraining, which undermines their efficiency. In this paper, we present a systematic analysis of language sensitivity by varying either the NAC training language or the SSL pre-training language while keeping the other fixed. Experimental results show that downstream performance is insensitive to the NAC training language but strongly dependent on the SSL pre-training language. These findings suggest that a single NAC can be reused across languages, while aligning the SSL pre-training language with the target language is crucial.

ESPnet3: Infrastructure for Scalable Speech and Audio Research in the Foundation Model Era

Recent speech research involves increasingly large datasets, complex models, and diverse experimental workflows. However, existing frameworks require substantial engineering effort to support such experiments. We present ESPnet3, a speech and audio research framework built on a modular system architecture with configuration-driven dataset composition and unified Python-based workflows. ESPnet3 introduces a DataOrganizer abstraction for flexible dataset integration and dataset sharding for memory-efficient large-scale training, while allowing recipe-specific logic through lightweight stage overrides. In OWSM pre-training experiments, ESPnet3 reduces per-epoch training time by \emph{21.1 minutes} compared to ESPnet2 and achieves \emph{>80\% GPU utilization} in multi-node training. Fine-tuning experiments show that new models and datasets can be integrated with around \emph{46 lines of additional code}. ESPnet3 will be publicly released with model checkpoints and training logs.

Grounding Spoken LLMs in Multi-Speaker Audio via Diarization Conditioning

We propose diarization-conditioned spoken language models (SLMs), a strategy for extending SLMs to far-field multi-talker audio. Rather than adapting the decoder via Serialized Output Training, which risks catastrophic forgetting, we condition the acoustic encoder on diarization masks to extract target-speaker representations, keeping the decoder frozen. We instantiate this as Dixtral, integrating a Diarization Conditioned Whisper (DiCoW) encoder into the Voxtral SLM. On AMI, NOTSOFAR-1, LibriSpeechMix, and Mixer6, Dixtral outperforms Gemini 3.0 Flash, VibeVoice, and Voxtral Mini Transcribe V2 on speaker-attributed transcription by 29.0%, 19.8%, and 16.0% absolute cpWER respectively. On a novel long-form multi-speaker QA benchmark, zero-shot Dixtral matches Gemini on far-field content understanding, and when fine-tuned surpasses both Gemini and Voxtral operating on close-talk across all tasks.

Exploiting Noise Inseparability for Weakly-Supervised Discriminative Speech Denoising Using Noisy Targets

Speech denoising is an often necessary step not only for human listening, but also for downstream processing by systems lacking robustness to noisy, real-world acoustic conditions. Unfortunately, denoising is a problem where conventional in-domain supervised training is not trivial, as the training targets cannot be annotated by humans: producing a clean version of a naturally-noisy speech recording is itself the task to solve. Supervised training is typically performed through the artificial addition of noise to clean speech recordings, which can only be sourced from controlled domains, a significant limitation due to the poor out-of-domain generalization of neural networks. An alternative is noisy target training (NyTT), which simply replaces the clean speech with in-domain noisy recordings, with the hope that learning to remove the artificial noise will extend to the natural. Though having shown promising results, NyTT's training objective is not minimized by clean speech estimates. We show that by estimating the artificial noise in addition to the naturally-noisy speech, the undesirable optimum can actually be exploited: the residual noise in the speech estimate can be canceled by the noise estimate via simple subtraction. Crucially, the optimum is fully compatible with conventional artificial mixtures, enabling joint training using both types of data with consistent optimization targets, opening the door to improved domain adaptability. The effectiveness of our approach is demonstrated through WHAM! and CHiME-3-based benchmarks.

Cross-Talk Speech Reduction, by Separation, for Separation

In conversational speech separation and recognition tasks, close-talk microphones are typically attached to each speaker during training data collection to capture near-field, close-talk mixture signals, in addition to using far-field microphones to record far-field mixture signals. Each such close-talk mixture exhibits a reasonably high energy level for the wearer and could intuitively serve as weak supervision for training far-field speech separation models directly on real-recorded far-field signals. However, they are not sufficiently clean for this purpose, as they often contain strong cross-talk speech from other speakers in addition to background noise. To address this, we propose cross-talk reduction (CTR), a task aiming to isolate the wearer's speech from each close-talk mixture, and a novel method called CTRnet, which can be trained directly on real-recorded pairs of close-talk and far-field mixtures to accomplish CTR. Building on CTRnet, we further propose pseudo-label based far-field speech separation (PuLSS), which uses CTRnet's estimated clean speech as pseudo-labels to train models for separating far-field mixtures. A key advantage of the proposed framework is that both CTRnet and PuLSS can be trained on real-recorded data from the target domain, addressing the generalization gap commonly observed when models are trained exclusively on simulated data. On the CHiME-6 dataset, our framework achieves state-of-the-art ASR performance under both oracle and estimated speaker diarization, surpassing all CHiME-{7,8} challenge submissions. To our knowledge, it is the first neural speech separation method that substantially outperforms guided source separation on real conversational "speech-in-the-wild" data.

Ring Mixing with Auxiliary Signal-to-Consistency-Error Ratio Loss for Unsupervised Denoising in Speech Separation

Noisy speech separation systems are typically trained on fully-synthetic mixtures, limiting generalization to real-world scenarios. Though training on mixtures of in-domain (thus often noisy) speech is possible, we show that this leads to undesirable optima where mixture noise is retained in the estimates, due to the inseparability of the background noises and the loss function's symmetry. To address this, we propose ring mixing, a batch strategy of using each source in two mixtures, alongside a new Signal-to-Consistency-Error Ratio (SCER) auxiliary loss penalizing inconsistent estimates of the same source from different mixtures, breaking symmetry and incentivizing denoising. On a WHAM!-based benchmark, our method can reduce residual noise by upwards of half, effectively learning to denoise from only noisy recordings. This opens the door to training more generalizable systems using in-the-wild data, which we demonstrate via systems trained using naturally-noisy speech from VoxCeleb.

Who Spoke What When? Evaluating Spoken Language Models for Conversational ASR with Semantic and Overlap-Aware Metrics

Conversational automatic speech recognition remains challenging due to overlapping speech, far-field noise, and varying speaker counts. While recent LLM-based systems perform well on single-speaker benchmarks, their robustness in multi-speaker settings is unclear. We systematically compare LLM-based and modular pipeline approaches along four axes: overlap robustness, semantic fidelity, speaker count, and single- versus multi-channel input. To capture meaning-altering errors that conventional metrics miss, we introduce tcpSemER, which extends tcpWER by replacing Levenshtein distance with embedding-based semantic similarity. We further decompose tcpWER into overlapping and non-overlapping components for finer-grained analysis. Experiments across three datasets show that LLM-based systems are competitive in two-speaker settings but degrade as speaker count and overlap increase, whereas modular pipelines remain more robust.

Modeling Overlapped Speech with Shuffles

We propose to model parallel streams of data, such as overlapped speech, using shuffles. Specifically, this paper shows how the shuffle product and partial order finite-state automata (FSAs) can be used for alignment and speaker-attributed transcription of overlapped speech. We train using the total score on these FSAs as a loss function, marginalizing over all possible serializations of overlapping sequences at subword, word, and phrase levels. To reduce graph size, we impose temporal constraints by constructing partial order FSAs. We address speaker attribution by modeling (token, speaker) tuples directly. Viterbi alignment through the shuffle product FSA directly enables one-pass alignment. We evaluate performance on synthetic LibriSpeech overlaps. To our knowledge, this is the first algorithm that enables single-pass alignment of multi-talker recordings. All algorithms are implemented using k2 / Icefall.

SE-DiCoW: Self-Enrolled Diarization-Conditioned Whisper

Speaker-attributed automatic speech recognition (ASR) in multi-speaker environments remains a major challenge. While some approaches achieve strong performance when fine-tuned on specific domains, few systems generalize well across out-of-domain datasets. Our prior work, Diarization-Conditioned Whisper (DiCoW), leverages speaker diarization outputs as conditioning information and, with minimal fine-tuning, demonstrated strong multilingual and multi-domain performance. In this paper, we address a key limitation of DiCoW: ambiguity in Silence-Target-Non-target-Overlap (STNO) masks, where two or more fully overlapping speakers may have nearly identical conditioning despite differing transcriptions. We introduce SE-DiCoW (Self-Enrolled Diarization-Conditioned Whisper), which uses diarization output to locate an enrollment segment anywhere in the conversation where the target speaker is most active. This enrollment segment is used as fixed conditioning via cross-attention at each encoder layer. We further refine DiCoW with improved data segmentation, model initialization, and augmentation. Together, these advances yield substantial gains: SE-DiCoW reduces macro-averaged tcpWER by 52.4% relative to the original DiCoW on the EMMA MT-ASR benchmark.

The CMU-AIST submission for the ICME 2025 Audio Encoder Challenge

This technical report describes our submission to the ICME 2025 audio encoder challenge. Our submitted system is built on BEATs, a masked speech token prediction based audio encoder. We extend the BEATs model using 74,000 hours of data derived from various speech, music, and sound corpora and scale its architecture upto 300 million parameters. We experiment with speech-heavy and balanced pre-training mixtures to study the impact of different domains on final performance. Our submitted system consists of an ensemble of the Dasheng 1.2 billion model with two custom scaled-up BEATs models trained on the aforementioned pre-training data mixtures. We also propose a simple ensembling technique that retains the best capabilities of constituent models and surpasses both the baseline and Dasheng 1.2B. For open science, we publicly release our trained checkpoints via huggingface at https://huggingface.co/shikhar7ssu/OpenBEATs-ICME-SOUND and https://huggingface.co/shikhar7ssu/OpenBEATs-ICME.