2026· Annual Meeting of the Association for Computational Linguistics· pp. 35363-35377· 0 citations· 36 references
Computer Science
TL;DR
Experiments on Europarl-ST show that LCMA-SRT improves both ASR and ST within a single joint model, reducing average WER and improving BLEU and COMET over strong hierarchical transducer baselines.
Abstract
Neural transducers offer an alignment-free framework for speech-to-text modeling, and hierarchical transducer architectures further improve multilingual joint automatic speech recognition (ASR) and speech translation (ST) by stacking a translation-focused encoder on top of an ASR encoder. However, extending hierarchical transducers to multilingual many-to-many settings remains challenging: fully shared models often suffer from negative transfer and unstable target-language generation, while training separate models for each direction is computationally prohibitive. We propose LCMA-SRT (Language-Conditional Mixture-of-Experts Adapters for Speech Recognition and Translation), which augments a hierarchical transducer with language-conditional Mixture-of-Experts (MoE) adapters. A source-conditioned MoE adapter (SRC-MoE) uses source-language embeddings to reduce cross-language interference and improve multilingual ASR. A target-conditioned MoE adapter (TGT-MoE) uses the desired target language to reduce cross-target interference and stabilize target-language generation in many-to-many ST. Experiments on Europarl-ST (9 languages, 72 directions) show that LCMA-SRT improves both ASR and ST within a single joint model, reducing average WER and improving BLEU and COMET over strong hierarchical transducer baselines. We release our code and models at https:
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.
Yexing Du, Kaiyuan Liu, Youcheng Pan et al.· 0 citations
MEUSLI is introduced, 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.
Lorenzo Concina, Seraphina Fong, M. Matassoni et al.· 0 citations
Target-speaker automatic speech recognition (TS-ASR) enables transcription of a specific speaker in multi-talker environments, yet remains largely unexplored for multilingual, low-resource languages. Existing TS-ASR systems predominantly target monolingual English using diarization-based or speaker-embedding approaches, leaving a critical gap for languages such as Kazakh, where code-switching with Russian and English is commonplace. We propose Persona-ASR, a modular two-stage architecture. The first stage is an explicit target-presence gate that verifies whether the enrolled speaker appears in the mixture and emits a token to suppress transcription when the speaker is absent, directly addressing the acoustic-hallucination failure mode of prior systems. The second stage performs enrollment-conditioned recognition: a 192-dimensional ECAPA-TDNN speaker embedding modulates a WavLM-Base-Plus encoder through feature-wise linear modulation (FiLM), while language-specific CTC heads enable joint Kazakh and English decoding without forcing Latin and Cyrillic symbols to compete in a single output space. To evaluate the system, we introduce KazMix3, a Kazakh overlap dataset for TS-ASR training, and PersonaMix, a controlled bilingual benchmark spanning same- and cross-language enrollment across varying interferer counts (1–3) and signal-to-noise ratios (−3 to +3 dB). Persona-ASR outperforms a strong off-the-shelf cascade baseline by 13.3 WER points on English and 24.6 on Kazakh, and matches a published monolingual English baseline. On PersonaMix, speaker conditioning reduces relative word error rate by 40.7% on English and 59.3% on Kazakh mixtures over an unconditioned variant of the same model, and cross-language enrollment (unseen during training) remains effective, increasing average raw WER by only 4.1 points (English) and 2.2 points (Kazakh) relative to same-language enrollment. To our knowledge, Persona-ASR is the first TS-ASR system for the Kazakh language, establishing a foundation for multilingual personalized ASR in low-resource settings.
Rakhat Meiramov, Tomiris Rakhimzhanova, Adil Taibassarov et al.· Machine Learning and Knowled...· 0 citations
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.
Huaxuan Wang, Huimin Wang, Ruiyu Zhang et al.· 0 citations
The results suggest that competitive MLLM can emerge from alignment alone, reducing multimodal extension to a lightweight projector-training problem that generalizes across modalities and adapts rapidly to each new LLM release.
Xuanru Zhou, Yiwen Shao, Jiahong Li et al.· 1 citation
Audio-encoder-LLM-decoder architectures have become the dominant paradigm for modern automatic speech recognition (ASR), improving transcription quality through large-scale language modeling. However, the cost of autoregressive decoding scales with decoder size, creating a fundamental trade-off between recognition quality and serving latency. We argue this trade-off is not inherent: unlike open-ended text generation, ASR outputs are strongly anchored to the input speech signal, providing a natural inductive bias toward high-parallelism decoding. Building on this, we introduce ParaASR, an ASR system that leverages Multi-Token Prediction (MTP) to let a 4B LLM decoder emit multiple tokens per forward step. Starting from a publicly available audio-language foundation, the model first establishes a robust autoregressive recognizer and then aligns five future-token branches through a staged optimization recipe. At inference, it proposes a six-token continuation per step and admits only the verified prefix into the transcript, preserving the safety of standard autoregressive decoding. The average accepted length reaches 5.0 out of 6 proposed tokens, confirming that the deterministic structure of speech makes ASR an especially natural setting for multi-token decoding. ParaASR further retains a native 32K-context window and transcribes up to 30 minutes of audio in a single pass. Across diverse benchmarks, it attains average error rates of 2.97%, 3.68%, and 3.70% on Chinese, English, and long-form evaluations, respectively, while reaching a real-time factor (RTF) as low as 0.0053. These results show that decoder scaling, low-latency inference, and long-context transcription need not be competing goals when future-token proposals are anchored by the acoustic signal and guarded by autoregressive verification.
Qingjian Lin, Yuxin Li, Haoyang Zhang et al.· 1 citation