Weihua Zheng, Zhengyuan Liu, Tanmoy Chakraborty, Weiwen Xu, Xiaoxue Gao, Bryan Chen Zhengyu Tan, Bowei Zou, Chang Liu, Yujia Hu, Xing Xie, Xiaoyuan Yi, Jing Yao, Chaojun Wang, Long Li, Rui Liu, Huiyao Liu, Koji Inoue, Ryuichi Sumida, Tatsuya Kawahara, Fan Xu, Lingyu Ye, Wei Tian, Dongjun Kim, Jimin Jung, Jaehyung Seo, Nadya Yuki Wangsajaya, Pham Minh Duc, Ojasva Saxena, Palash Nandi, Xiyan Tao, Wiwik Karlina, Tuan Luong, Keertana Arun Vasan, Roy Ka-Wei Lee, Nancy F. Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Abstract Recent advancements in large language models (LLMs) like GPT-4o have enabled real-time speech interactions through LLM-based voice assistants, offering an improved user experience over text-based interactions. However, a suitable benchmark to rigorously evaluate such speech interactions systems is currently lacking. To bridge this gap, we introduce VoiceBench, the first benchmark specifically designed to assess LLM-based voice assistants. VoiceBench comprises 6,783 synthetic and real spoken instructions recorded from diverse speakers across eight distinct tasks. These instructions are meticulously crafted to assess three crucial capability areas: general knowledge, instruction-following, and safety compliance. Furthermore, VoiceBench systematically incorporates realistic variations common in spoken interactions, including differences in speaker characteristics (e.g., accents), heterogeneous environmental conditions (e.g., reverberation), and content complexities such as mispronunciations. Extensive experiments reveal the limitations of current LLM-based voice assistant models and offer valuable insights for future research and development in this field.1

Automatic Speech Recognition (ASR) systems are pivotal in transcribing speech into text, yet the errors they introduce can significantly degrade the performance of downstream tasks like summarization. This issue is particularly pronounced in clinical dialogue summarization, a low-resource domain where supervised data for fine-tuning is scarce, necessitating the use of ASR models as black-box solutions. Employing conventional data augmentation for enhancing the noise robustness of summarization models is not feasible either due to the unavailability of sufficient medical dialogue audio recordings and corresponding ASR transcripts. To address this challenge, we propose MEDSAGE, an approach for generating synthetic samples for data augmentation using Large Language Models (LLMs). Specifically, we leverage the in-context learning capabilities of LLMs and instruct them to generate ASR-like errors based on a few available medical dialogue examples with audio recordings. Experimental results show that LLMs can effectively model ASR noise, and incorporating this noisy data into the training process significantly improves the robustness and accuracy of medical dialogue summarization systems. This approach addresses the challenges of noisy ASR outputs in critical applications, offering a robust solution to enhance the reliability of clinical dialogue summarization.

Voiced Electromyography(EMG)-to-Speech (V-ETS) models reconstruct speech from muscle activity signals, facilitating applications such as neurolaryngologic diagnostics. Despite its potential, the advancement of V-ETS is hindered by a scarcity of paired EMG-speech data. To address this, we propose a novel Confidence-based Multi-Speaker Self-training (CoM2S) approach, along with a newly curated Libri-EMG dataset. This approach leverages synthetic EMG data generated by a pretrained model, followed by a proposed filtering mechanism based on phoneme-level confidence to enhance the V-ETS model through the proposed self-training techniques. Experiments demonstrate our method improves phoneme accuracy, reduces phonological confusion, and lowers word error rate, confirming the effectiveness of our CoM2S approach for V-ETS. In support of future research, we will release the codes and the proposed Libri-EMG dataset-an open-access, time-aligned, multi-speaker voiced EMG and speech recordings.

Current emotional text-to-speech (TTS) models pre-dominantly conduct supervised training to learn the conversion from text and desired emotion to its emotional speech, focusing on a single emotion per text-speech pair. These models only learn the correct emotional outputs without fully comprehending other emotion characteristics, which limits their capabilities of capturing the nuances between different emotions. We propose a controllable Emo-DPO approach, which employs direct preference optimization to differentiate subtle emotional nuances between emotions through optimizing towards preferred emotions over less preferred emotional ones. Instead of relying on traditional neural architectures used in existing emotional TTS models, we propose utilizing the emotion-aware LLM-TTS neural architecture to leverage LLMs’ in-context learning and instruction-following capabilities. Comprehensive experiments confirm that our proposed method outperforms the existing baselines.

Existing expressive text-to-speech (TTS) systems primarily model a limited set of categorical emotions, whereas human conversations extend far beyond these predefined emotions, making it essential to explore more diverse emotional speech generation for more natural interactions. To bridge this gap, this paper proposes a novel prompt-unseen-emotion (PUE) approach to generate unseen emotional speech via emotion-guided prompt learning.PUEis trained utilizing an LLM-TTS architecture to ensure emotional consistency between categorical emotion-relevant prompts and emotional speech, allowing the model to quantitatively capture different emotion weightings per utterance. During inference, mixed emotional speech can be generated by flexibly adjusting emotion proportions and leveraging LLM contextual knowledge, enabling the model to quantify different emotional styles. Our proposedPUEsuccessfully facilitates expressive speech synthesis of unseen emotions in a zero-shot setting.

Audio-visual active speaker detection (AV-ASD) aims to identify which visible face is speaking in a scene with one or more persons. Most existing AV-ASD methods prioritize capturing speech-lip correspondence. However, there is a noticeable gap in addressing the challenges from real-world AV-ASD scenarios. Due to the presence of low-quality noisy videos in such cases, AV-ASD systems without a selective listening ability are short of effectively filtering out disruptive voice components from mixed audio inputs. In this paper, we propose a Multi-modal Speech Extraction-to-Detection framework named ‘MuSED’, which is pre-trained with audio-visual target speech extraction to learn the denoising ability, then it is fine-tuned with the AV-ASD task. Meanwhile, to better capture the multi-modal information and deal with real-world problems such as missing modality, MuSED is modelled on the time domain directly and integrates the multi-modal plus-and-minus augmentation strategy. Our experiments demonstrate that MuSED substantially outperforms the state-of-the-art AV-ASD methods and achieves 95.6% mAP on the AVA-ActiveSpeaker dataset, 98.3% AP on the ASW dataset, and 97.9% F1 on the Columbia AV-ASD dataset, respectively. We will publicly release the code in due course.

Current automatic speech recognition systems struggle with modeling long speech sequences due to high quadratic complexity of Transformer-based models. Selective state space models such as Mamba has performed well on long-sequence modeling in natural language processing and computer vision tasks. However, research endeavors in speech technology tasks has been under-explored. We propose Speech-Mamba, which incorporates selective state space modeling in Transformer neural architectures. Long sequence representations with selective state space models in Speech-Mamba is complemented with lower-level representations from Transformer-based modeling. Speech-mamba achieves better capacity to model long-range dependencies, as it scales near-linearly with sequence length.

Given the extensive research and real-world applications of automatic speech recognition (ASR), ensuring the robustness of ASR models against minor input perturbations becomes a crucial consideration for maintaining their effectiveness in real-time scenarios. Previous explorations into ASR model robustness have predominantly revolved around evaluating accuracy on white-box settings with full access to ASR models. Nevertheless, full ASR model details are often not available in real-world applications. Therefore, evaluating the robustness of black-box ASR models is essential for a comprehensive understanding of ASR model resilience. In this regard, we thoroughly study the vulnerability of practical black-box attacks in cutting-edge ASR models and propose to employ two advanced time-domain-based transferable attacks alongside our differentiable feature extractor. We also propose a speech-aware gradient optimization approach (SAGO) for ASR, which forces mistranscription with minimal impact on human imperceptibility through voice activity detection rule and a speech-aware gradient-oriented optimizer. Our comprehensive experimental results reveal performance enhancements compared to baseline approaches across five models on two databases.

Human language can be expressed in either written or spoken form, i.e. text or speech. Humans can acquire knowledge from text to improve speaking and listening. However, the quest for speech pre-trained models to leverage unpaired text has just started. In this letter, we investigate a new way to pre-train such a joint speech-text model to learn enhanced speech representations and benefit various speech-related downstream tasks. Specifically, we propose a novel pre-training method, text-guided HuBERT, or T-HuBERT, which performs self-supervised learning over speech to derive phoneme-like discrete representations. And these phoneme-like pseudo-label sequences are firstly derived from speech via the generative adversarial networks (GAN) to be statistically similar to those from additional unpaired textual data. In this way, we build a bridge between unpaired speech and text in a unsupervised manner. Extensive experiments demonstrate the significant superiority of our proposed method over various strong baselines, which achieves up to 15.3% relative Word Error Rate (WER) reduction on the LibriSpeech dataset.

The current lyrics transcription approaches heavily rely on supervised learning with labeled data, but such data are scarce and manual labeling of singing is expensive. How to benefit from unlabeled data and alleviate limited data problem have not been explored for lyrics transcription. We propose the first semi-supervised lyrics transcription paradigm, Self-Transcriber, by leveraging on unlabeled data using selftraining with noisy student augmentation. We attempt to demonstrate the possibility of lyrics transcription with a few amount of labeled data. Self-Transcriber generates pseudo labels of the unlabeled singing using teacher model, and augments pseudo-labels to the labeled data for student model update with both self-training and supervised training losses. This work closes the gap between supervised and semi- supervised learning as well as opens doors for few-shot learning of lyrics transcription. Our experiments show that our approach using only 12.7 hours of labeled data achieves competitive performance compared with the supervised approaches trained on 149.1 hours of labeled data for lyrics transcription.

Self-supervised pre-training has been successful in both text and speech processing. Speech and text offer different but complementary information. The question is whether we are able to perform a speech-text joint pre-training on unpaired speech and text. In this paper, we take the idea of self-supervised pre-training one step further and propose token2vec, a novel joint pre-training framework for unpaired speech and text based on discrete representations of speech. Specifically, we introduce two modality-specific tokenizers for speech and text. Based on these tokenizers, we convert speech/text sequences into discrete speech/text token sequences consisting of similar language units, thus mitigating the domain mismatch problem and length mismatch problem, which are caused by the distinct characteristics between speech and text. Finally, we feed the discrete speech and text tokens into a modality-agnostic Transformer encoder and pre-train with token-level masking language modeling (tMLM). Experiments show that token2vec is significantly superior to various speech-only pre-training baselines, with up to 17.7% relative WER reduction. Token2vec model is also validated on a non-ASR task, i.e., spoken intent classification, and shows good transferability.

Lyrics transcription of polyphonic music is challenging as the background music affects lyrics intelligibility. Typically, lyrics transcription can be performed by a two-step pipeline, i.e. a singing vocal extraction front end, followed by a lyrics transcriber back end, where the front end and back end are trained separately. Such a two-step pipeline suffers from both imperfect vocal extraction and mismatch between front end and back end. In this work, we propose a novel end-to-end integrated fine-tuning framework, that we call PoLyScriber, to globally optimize the vocal extractor front end and lyrics transcriber back end for lyrics transcription in polyphonic music. The experimental results show that our proposed PoLyScriber achieves substantial improvements over the existing approaches on publicly available test datasets.

Lyrics transcription of polyphonic music is challenging not only because the singing vocals are corrupted by the background music, but also because the background music and the singing style vary across music genres, such as pop, metal, and hip hop, which affects lyrics intelligibility of the song in different ways. In this work, we propose to transcribe the lyrics of polyphonic music using a novel genre-conditioned network. The proposed network adopts pre-trained model parameters, and incorporates the genre adapters between layers to capture different genre peculiarities for lyrics-genre pairs, thereby only requiring lightweight genre-specific parameters for training. Our experiments show that the proposed genre-conditioned network outperforms the existing lyrics transcription systems.

Lyrics are the words that make up a song, while chords are harmonic sets of multiple notes in music. Lyrics and chords are generally essential information in music, i.e. unaccompanied singing vocals mixed with instrumental music, representing important components in polyphonic music. In a traditional lyrics transcription task, we first extract the singing vocals from the polyphonic music and then transcribe the resulting singing vocals, where the two steps are optimized independently. In this paper, we propose novel end-to-end network architectures that are designed to disentangle lyrics from chords in polyphonic music for effective lyrics transcription in a single step, where we consider chords as musical words, analogously to lexical words as lyrics intuitively. We start by studying a single-task lyrics transcriber as the reference baseline and the initial model to develop the multi-task lyrics transcription solutions. The main idea is to take advantage of chord transcription available in the training data through multi-task training to improve lyrics transcription. The experiments show that the proposed multitask lyrics transcriber significantly outperforms other competing solutions, with a word error rate (WER) of 31.82% on a standard test dataset.

Design efficiency can be improved by timely pushing knowledge. Knowledge is needed when designer meets difficulty. If the features that reflect a designer encounters difficulty are captured and identified, a time slot can be decided to push related knowledge. The behaviour pattern when designer encounters difficulty is thus investigated, which will facilitate identifying whether designer meets difficulty or not. First, design behaviour and behaviour rules are acquired through behaviour observation experiment. And then in order to systematically describe behaviour, Behaviour hierarchy model is proposed based on elements and rules of behaviour. Finally, based on high-level behaviour of hierarchy model, behaviour pattern with emotion-state as label is summarized and verified by experiment data. The work results describe behaviour formally and provide behaviour pattern as a basis for identifying whether designers meet with difficulty.