Multi-Distillation from Speech and Music Representation Models

Real-world audio often mixes speech and music, yet models typically handle only one domain. This paper introduces a multi-teacher distillation framework that unifies speech and music models into a single one while significantly reducing model size. Our approach leverages the strengths of domain-specific teacher models, such as HuBERT for speech and MERT for music, and explores various strategies to balance both domains. Experiments across diverse tasks demonstrate that our model matches the performance of domain-specific models, showing the effectiveness of cross-domain distillation. Additionally, we conduct few-shot learning experiments, highlighting the need for general models in real-world scenarios where labeled data is limited. Our results show that our model not only performs on par with specialized models but also outperforms them in few-shot scenarios, proving that a cross-domain approach is essential and effective for diverse tasks with limited data.

E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models

Speech Foundation Models encounter significant performance degradation when deployed in real-world scenarios involving acoustic domain shifts, such as background noise and speaker accents. Test-time adaptation (TTA) has recently emerged as a viable strategy to address such domain shifts at inference time without requiring access to source data or labels. However, existing TTA approaches, particularly those relying on backpropagation, are memory-intensive, limiting their applicability in speech tasks and resource-constrained settings. Although backpropagation-free methods offer improved efficiency, existing ones exhibit poor accuracy. This is because they are predominantly developed for vision tasks, which fundamentally differ from speech task formulations, noise characteristics, and model architecture, posing unique transferability challenges. In this paper, we introduce E-BATS, the first Efficient BAckpropagation-free TTA framework designed explicitly for speech foundation models. E-BATS achieves a balance between adaptation effectiveness and memory efficiency through three key components: (i) lightweight prompt adaptation for a forward-pass-based feature alignment, (ii) a multi-scale loss to capture both global (utterance-level) and local distribution shifts (token-level) and (iii) a test-time exponential moving average mechanism for stable adaptation across utterances. Experiments conducted on four noisy speech datasets spanning sixteen acoustic conditions demonstrate consistent improvements, with 4.1%-13.5% accuracy gains over backpropagation-free baselines and 2.0-6.4 times GPU memory savings compared to backpropagation-based methods. By enabling scalable and robust adaptation under acoustic variability, this work paves the way for developing more efficient adaptation approaches for practical speech processing systems in real-world environments.

Speech Recognition on TV Series with Video-guided Post-Correction

Automatic Speech Recognition (ASR) has achieved remarkable success with deep learning, driving advancements in conversational artificial intelligence, media transcription, and assistive technologies. However, ASR systems still struggle in complex environments such as TV series, where overlapping speech, domain-specific terminology, and long-range contextual dependencies pose significant challenges to transcription accuracy. Existing multimodal approaches fail to correct ASR outputs with the rich temporal and contextual information available in video. To address this limitation, we propose a novel multimodal post-correction framework that refines ASR transcriptions by leveraging contextual cues extracted from video. Our framework consists of two stages: ASR Generation and Video-based Post-Correction, where the first stage produces the initial transcript and the second stage corrects errors using Video-based Contextual Information Extraction and Context-aware ASR Correction. We employ the Video-Large Multimodal Model (VLMM) to extract key contextual information using tailored prompts, which is then integrated with a Large Language Model (LLM) to refine the ASR output. We evaluate our method on a multimodal benchmark for TV series ASR and demonstrate its effectiveness in improving ASR performance by leveraging video-based context to enhance transcription accuracy in complex multimedia environments.

Streaming Endpointer for Spoken Dialogue using Neural Audio Codecs and Label-Delayed Training

Accurate, low-latency endpointing is crucial for effective spoken dialogue systems. While traditional endpointers often rely on spectrum-based audio features, this work proposes real-time speech endpointing for multi-turn dialogues using streaming, low-bitrate Neural Audio Codec (NAC) features, building upon recent advancements in neural audio codecs. To further reduce cutoff errors, we introduce a novel label delay training scheme. At a fixed median latency of 160 ms, our combined NAC and label delay approach achieves significant relative cutoff error reductions: 42.7% for a single-stream endpointer and 37.5% for a two-stream configuration, compared to baseline methods. Finally, we demonstrate efficient integration with a codec-based pretrained speech large language model, improving its median response time by 1200 ms and reducing its cutoff error by 35%.

"In This Environment, As That Speaker": A Text-Driven Framework for Multi-Attribute Speech Conversion

We propose TES-VC (Text-driven Environment and Speaker controllable Voice Conversion), a text-driven voice conversion framework with independent control of speaker timbre and environmental acoustics. TES-VC processes simultaneous text inputs for target voice and environment, accurately generating speech matching described timbre/environment while preserving source content. Trained on synthetic data with decoupled vocal/environment features via latent diffusion modeling, our method eliminates interference between attributes. The Retrieval-Based Timbre Control (RBTC) module enables precise manipulation using abstract descriptions without paired data. Experiments confirm TES-VC effectively generates contextually appropriate speech in both timbre and environment with high content retention and superior controllability which demonstrates its potential for widespread applications.

Technical Report: A Practical Guide to Kaldi ASR Optimization

This technical report introduces innovative optimizations for Kaldi-based Automatic Speech Recognition (ASR) systems, focusing on acoustic model enhancement, hyperparameter tuning, and language model efficiency. We developed a custom Conformer block integrated with a multistream TDNN-F structure, enabling superior feature extraction and temporal modeling. Our approach includes advanced data augmentation techniques and dynamic hyperparameter optimization to boost performance and reduce overfitting. Additionally, we propose robust strategies for language model management, employing Bayesian optimization and $n$-gram pruning to ensure relevance and computational efficiency. These systematic improvements significantly elevate ASR accuracy and robustness, outperforming existing methods and offering a scalable solution for diverse speech recognition scenarios. This report underscores the importance of strategic optimizations in maintaining Kaldi's adaptability and competitiveness in rapidly evolving technological landscapes.

Rhythm Features for Speaker Identification

While deep learning models have demonstrated robust performance in speaker recognition tasks, they primarily rely on low-level audio features learned empirically from spectrograms or raw waveforms. However, prior work has indicated that idiosyncratic speaking styles heavily influence the temporal structure of linguistic units in speech signals (rhythm). This makes rhythm a strong yet largely overlooked candidate for a speech identity feature. In this paper, we test this hypothesis by applying deep learning methods to perform text-independent speaker identification from rhythm features. Our findings support the usefulness of rhythmic information for speaker recognition tasks but also suggest that high intra-subject variability in ad-hoc speech can degrade its effectiveness.

SynHate: Detecting Hate Speech in Synthetic Deepfake Audio

The rise of deepfake audio and hate speech, powered by advanced text-to-speech, threatens online safety. We present SynHate, the first multilingual dataset for detecting hate speech in synthetic audio, spanning 37 languages. SynHate uses a novel four-class scheme: Real-normal, Real-hate, Fake-normal, and Fake-hate. Built from MuTox and ADIMA datasets, it captures diverse hate speech patterns globally and in India. We evaluate five leading self-supervised models (Whisper-small/medium, XLS-R, AST, mHuBERT), finding notable performance differences by language, with Whisper-small performing best overall. Cross-dataset generalization remains a challenge. By releasing SynHate and baseline code, we aim to advance robust, culturally sensitive, and multilingual solutions against synthetic hate speech. The dataset is available at https://www.iab-rubric.org/resources.

Accurate analysis of the pitch pulse-based magnitude/phase structure of natural vowels and assessment of three lightweight time/frequency voicing restoration methods

Whispered speech is produced when the vocal folds are not used, either intentionally, or due to a temporary or permanent voice condition. The essential difference between natural speech and whispered speech is that periodic signal components that exist in certain regions of the former, called voiced regions, as a consequence of the vibration of the vocal folds, are missing in the latter. The restoration of natural speech from whispered speech requires delicate signal processing procedures that are especially useful if they can be implemented on low-resourced portable devices, in real-time, and on-the-fly, taking advantage of the established source-filter paradigm of voice production and related models. This paper addresses two challenges that are intertwined and are key in informing and making viable this envisioned technological realization. The first challenge involves characterizing and modeling the evolution of the harmonic phase/magnitude structure of a sequence of individual pitch periods in a voiced region of natural speech comprising sustained or co-articulated vowels. This paper proposes a novel algorithm segmenting individual pitch pulses, which is then used to obtain illustrative results highlighting important differences between sustained and co-articulated vowels, and suggesting practical synthetic voicing approaches. The second challenge involves model-based synthetic voicing. Three implementation alternatives are described that differ in their signal reconstruction approaches: frequency-domain, combined frequency and time-domain, and physiologically-inspired separate filtering of glottal excitation pulses individually generated. The three alternatives are compared objectively using illustrative examples, and subjectively using the results of listening tests involving synthetic voicing of sustained and co-articulated vowels in word context.

Automatic Speech Recognition of African American English: Lexical and Contextual Effects

Automatic Speech Recognition (ASR) models often struggle with the phonetic, phonological, and morphosyntactic features found in African American English (AAE). This study focuses on two key AAE variables: Consonant Cluster Reduction (CCR) and ING-reduction. It examines whether the presence of CCR and ING-reduction increases ASR misrecognition. Subsequently, it investigates whether end-to-end ASR systems without an external Language Model (LM) are more influenced by lexical neighborhood effect and less by contextual predictability compared to systems with an LM. The Corpus of Regional African American Language (CORAAL) was transcribed using wav2vec 2.0 with and without an LM. CCR and ING-reduction were detected using the Montreal Forced Aligner (MFA) with pronunciation expansion. The analysis reveals a small but significant effect of CCR and ING on Word Error Rate (WER) and indicates a stronger presence of lexical neighborhood effect in ASR systems without LMs.