Chongjia Ni

Keyu An, Qian Chen, Chong Deng et al.

This report introduces FunAudioLLM, a model family designed to enhance natural voice interactions between humans and large language models (LLMs). At its core are two innovative models: SenseVoice, which handles multilingual speech recognition, emotion recognition, and audio event detection; and CosyVoice, which facilitates natural speech generation with control over multiple languages, timbre, speaking style, and speaker identity. SenseVoice-Small delivers exceptionally low-latency ASR for 5 languages, and SenseVoice-Large supports high-precision ASR for over 50 languages, while CosyVoice excels in multi-lingual voice generation, zero-shot in-context learning, cross-lingual voice cloning, and instruction-following capabilities. The models related to SenseVoice and CosyVoice have been open-sourced on Modelscope and Huggingface, along with the corresponding training, inference, and fine-tuning codes released on GitHub. By integrating these models with LLMs, FunAudioLLM enables applications such as speech-to-speech translation, emotional voice chat, interactive podcasts, and expressive audiobook narration, thereby pushing the boundaries of voice interaction technology. Demos are available at https://fun-audio-llm.github.io, and the code can be accessed at https://github.com/FunAudioLLM.

Lei Wang, Rong Tong, Cheung Chi Leung et al.

This paper provides an overall introduction of our Automatic Speech Recognition (ASR) systems for Southeast Asian languages. As not much existing work has been carried out on such regional languages, a few difficulties should be addressed before building the systems: limitation on speech and text resources, lack of linguistic knowledge, etc. This work takes Bahasa Indonesia and Thai as examples to illustrate the strategies of collecting various resources required for building ASR systems.

Chong Zhang, Yukun Ma, Qian Chen et al.

We introduce InspireMusic, a framework integrated super resolution and large language model for high-fidelity long-form music generation. A unified framework generates high-fidelity music, songs, and audio, which incorporates an autoregressive transformer with a super-resolution flow-matching model. This framework enables the controllable generation of high-fidelity long-form music at a higher sampling rate from both text and audio prompts. Our model differs from previous approaches, as we utilize an audio tokenizer with one codebook that contains richer semantic information, thereby reducing training costs and enhancing efficiency. This combination enables us to achieve high-quality audio generation with long-form coherence of up to $8$ minutes. Then, an autoregressive transformer model based on Qwen 2.5 predicts audio tokens. Next, we employ a super-resolution flow-matching model to generate high-sampling rate audio with fine-grained details learned from an acoustic codec model. Comprehensive experiments show that the InspireMusic-1.5B-Long model has a comparable performance to recent top-tier open-source systems, including MusicGen and Stable Audio 2.0, on subjective and objective evaluations. The code and pre-trained models are released at https://github.com/FunAudioLLM/InspireMusic.

Yizhou Peng, Yi-Wen Chao, Dianwen Ng et al.

Full-duplex spoken dialogue systems (FDSDS) enable more natural human-machine interactions by allowing real-time user interruptions and backchanneling, compared to traditional SDS that rely on turn-taking. However, existing benchmarks lack metrics for FD scenes, e.g., evaluating model performance during user interruptions. In this paper, we present a comprehensive FD benchmarking pipeline utilizing LLMs, TTS, and ASR to address this gap. It assesses FDSDS's ability to handle user interruptions, manage delays, and maintain robustness in challenging scenarios with diverse novel metrics. We applied our benchmark to three open-source FDSDS (Moshi, Freeze-omni, and VITA-1.5) using over 40 hours of generated speech, with 293 simulated conversations and 1,200 interruptions. The results show that all models continue to face challenges, such as failing to respond to user interruptions, under frequent disruptions and noisy conditions. Demonstrations, data, and code will be released.

Keyu An, Yanni Chen, Chong Deng et al.

In recent years, automatic speech recognition (ASR) has witnessed transformative advancements driven by three complementary paradigms: data scaling, model size scaling, and deep integration with large language models (LLMs). However, LLMs are prone to hallucination, which can significantly degrade user experience in real-world ASR applications. In this paper, we present Fun-ASR, a large-scale, LLM-based ASR system that synergistically combines massive data, large model capacity, LLM integration, and reinforcement learning to achieve state-of-the-art performance across diverse and complex speech recognition scenarios. Moreover, Fun-ASR is specifically optimized for practical deployment, with enhancements in streaming capability, noise robustness, code-switching, hotword customization, and satisfying other real-world application requirements. Experimental results show that while most LLM-based ASR systems achieve strong performance on open-source benchmarks, they often underperform on real industry evaluation sets. Thanks to production-oriented optimizations, Fun-ASR achieves state-of-the-art performance on real application datasets, demonstrating its effectiveness and robustness in practical settings.

Qian Chen, Yafeng Chen, Yanni Chen et al.

Recent advancements in large language models (LLMs) and multimodal speech-text models have laid the groundwork for seamless voice interactions, enabling real-time, natural, and human-like conversations. Previous models for voice interactions are categorized as native and aligned. Native models integrate speech and text processing in one framework but struggle with issues like differing sequence lengths and insufficient pre-training. Aligned models maintain text LLM capabilities but are often limited by small datasets and a narrow focus on speech tasks. In this work, we introduce MinMo, a Multimodal Large Language Model with approximately 8B parameters for seamless voice interaction. We address the main limitations of prior aligned multimodal models. We train MinMo through multiple stages of speech-to-text alignment, text-to-speech alignment, speech-to-speech alignment, and duplex interaction alignment, on 1.4 million hours of diverse speech data and a broad range of speech tasks. After the multi-stage training, MinMo achieves state-of-the-art performance across various benchmarks for voice comprehension and generation while maintaining the capabilities of text LLMs, and also facilitates full-duplex conversation, that is, simultaneous two-way communication between the user and the system. Moreover, we propose a novel and simple voice decoder that outperforms prior models in voice generation. The enhanced instruction-following capabilities of MinMo supports controlling speech generation based on user instructions, with various nuances including emotions, dialects, and speaking rates, and mimicking specific voices. For MinMo, the speech-to-text latency is approximately 100ms, full-duplex latency is approximately 600ms in theory and 800ms in practice. The MinMo project web page is https://funaudiollm.github.io/minmo, and the code and models will be released soon.

Zhihao Du, Changfeng Gao, Yuxuan Wang et al.

In our prior works, we introduced a scalable streaming speech synthesis model, CosyVoice 2, which integrates a large language model (LLM) and a chunk-aware flow matching (FM) model, and achieves low-latency bi-streaming speech synthesis and human-parity quality. Despite these advancements, CosyVoice 2 exhibits limitations in language coverage, domain diversity, data volume, text formats, and post-training techniques. In this paper, we present CosyVoice 3, an improved model designed for zero-shot multilingual speech synthesis in the wild, surpassing its predecessor in content consistency, speaker similarity, and prosody naturalness. Key features of CosyVoice 3 include: 1) A novel speech tokenizer to improve prosody naturalness, developed via supervised multi-task training, including automatic speech recognition, speech emotion recognition, language identification, audio event detection, and speaker analysis. 2) A new differentiable reward model for post-training applicable not only to CosyVoice 3 but also to other LLM-based speech synthesis models. 3) Dataset Size Scaling: Training data is expanded from ten thousand hours to one million hours, encompassing 9 languages and 18 Chinese dialects across various domains and text formats. 4) Model Size Scaling: Model parameters are increased from 0.5 billion to 1.5 billion, resulting in enhanced performance on our multilingual benchmark due to the larger model capacity. These advancements contribute significantly to the progress of speech synthesis in the wild. We encourage readers to listen to the demo at https://funaudiollm.github.io/cosyvoice3.

Yizhou Peng, Yukun Ma, Chong Zhang et al.

While Text-to-Speech (TTS) systems can achieve fine-grained control over emotional expression via natural language prompts, a significant challenge emerges when the desired emotion (style prompt) conflicts with the semantic content of the text. This mismatch often results in unnatural-sounding speech, undermining the goal of achieving fine-grained emotional control. Classifier-Free Guidance (CFG) is a key technique for enhancing prompt alignment; however, its application to auto-regressive (AR) TTS models remains underexplored, which can lead to degraded audio quality. This paper directly addresses the challenge of style-content mismatch in AR TTS models by proposing an adaptive CFG scheme that adjusts to different levels of the detected mismatch, as measured using large language models or natural language inference models. This solution is based on a comprehensive analysis of CFG's impact on emotional expressiveness in state-of-the-art AR TTS models. Our results demonstrate that the proposed adaptive CFG scheme improves the emotional expressiveness of the AR TTS model while maintaining audio quality and intelligibility.

Kun Zhou, Shengkui Zhao, Yukun Ma et al.

Recent language model-based text-to-speech (TTS) frameworks demonstrate scalability and in-context learning capabilities. However, they suffer from robustness issues due to the accumulation of errors in speech unit predictions during autoregressive language modeling. In this paper, we propose a phonetic enhanced language modeling method to improve the performance of TTS models. We leverage self-supervised representations that are phonetically rich as the training target for the autoregressive language model. Subsequently, a non-autoregressive model is employed to predict discrete acoustic codecs that contain fine-grained acoustic details. The TTS model focuses solely on linguistic modeling during autoregressive training, thereby reducing the error propagation that occurs in non-autoregressive training. Both objective and subjective evaluations validate the effectiveness of our proposed method.

Jia Qi Yip, Tuan Truong, Dianwen Ng et al.

In this paper, we propose ACA-Net, a lightweight, global context-aware speaker embedding extractor for Speaker Verification (SV) that improves upon existing work by using Asymmetric Cross Attention (ACA) to replace temporal pooling. ACA is able to distill large, variable-length sequences into small, fixed-sized latents by attending a small query to large key and value matrices. In ACA-Net, we build a Multi-Layer Aggregation (MLA) block using ACA to generate fixed-sized identity vectors from variable-length inputs. Through global attention, ACA-Net acts as an efficient global feature extractor that adapts to temporal variability unlike existing SV models that apply a fixed function for pooling over the temporal dimension which may obscure information about the signal's non-stationary temporal variability. Our experiments on the WSJ0-1talker show ACA-Net outperforms a strong baseline by 5\% relative improvement in EER using only 1/5 of the parameters.

Yingzhu Zhao, Chongjia Ni, Cheung-Chi Leung et al.

Transformer models have been used in automatic speech recognition (ASR) successfully and yields state-of-the-art results. However, its performance is still affected by speaker mismatch between training and test data. Further finetuning a trained model with target speaker data is the most natural approach for adaptation, but it takes a lot of compute and may cause catastrophic forgetting to the existing speakers. In this work, we propose a unified speaker adaptation approach consisting of feature adaptation and model adaptation. For feature adaptation, we employ a speaker-aware persistent memory model which generalizes better to unseen test speakers by making use of speaker i-vectors to form a persistent memory. For model adaptation, we use a novel gradual pruning method to adapt to target speakers without changing the model architecture, which to the best of our knowledge, has never been explored in ASR. Specifically, we gradually prune less contributing parameters on model encoder to a certain sparsity level, and use the pruned parameters for adaptation, while freezing the unpruned parameters to keep the original model performance. We conduct experiments on the Librispeech dataset. Our proposed approach brings relative 2.74-6.52% word error rate (WER) reduction on general speaker adaptation. On target speaker adaptation, our method outperforms the baseline with up to 20.58% relative WER reduction, and surpasses the finetuning method by up to relative 2.54%. Besides, with extremely low-resource adaptation data (e.g., 1 utterance), our method could improve the WER by relative 6.53% with only a few epochs of training.

Zhiping Zeng, Van Tung Pham, Haihua Xu et al.

In this work, we study leveraging extra text data to improve low-resource end-to-end ASR under cross-lingual transfer learning setting. To this end, we extend our prior work [1], and propose a hybrid Transformer-LSTM based architecture. This architecture not only takes advantage of the highly effective encoding capacity of the Transformer network but also benefits from extra text data due to the LSTM-based independent language model network. We conduct experiments on our in-house Malay corpus which contains limited labeled data and a large amount of extra text. Results show that the proposed architecture outperforms the previous LSTM-based architecture [1] by 24.2% relative word error rate (WER) when both are trained using limited labeled data. Starting from this, we obtain further 25.4% relative WER reduction by transfer learning from another resource-rich language. Moreover, we obtain additional 13.6% relative WER reduction by boosting the LSTM decoder of the transferred model with the extra text data. Overall, our best model outperforms the vanilla Transformer ASR by 11.9% relative WER. Last but not least, the proposed hybrid architecture offers much faster inference compared to both LSTM and Transformer architectures.

Van Tung Pham, Haihua Xu, Yerbolat Khassanov et al.

The attention-based end-to-end (E2E) automatic speech recognition (ASR) architecture allows for joint optimization of acoustic and language models within a single network. However, in a vanilla E2E ASR architecture, the decoder sub-network (subnet), which incorporates the role of the language model (LM), is conditioned on the encoder output. This means that the acoustic encoder and the language model are entangled that doesn't allow language model to be trained separately from external text data. To address this problem, in this work, we propose a new architecture that separates the decoder subnet from the encoder output. In this way, the decoupled subnet becomes an independently trainable LM subnet, which can easily be updated using the external text data. We study two strategies for updating the new architecture. Experimental results show that, 1) the independent LM architecture benefits from external text data, achieving 9.3% and 22.8% relative character and word error rate reduction on Mandarin HKUST and English NSC datasets respectively; 2)the proposed architecture works well with external LM and can be generalized to different amount of labelled data.

Yerbolat Khassanov, Haihua Xu, Van Tung Pham et al.

The lack of code-switch training data is one of the major concerns in the development of end-to-end code-switching automatic speech recognition (ASR) models. In this work, we propose a method to train an improved end-to-end code-switching ASR using only monolingual data. Our method encourages the distributions of output token embeddings of monolingual languages to be similar, and hence, promotes the ASR model to easily code-switch between languages. Specifically, we propose to use Jensen-Shannon divergence and cosine distance based constraints. The former will enforce output embeddings of monolingual languages to possess similar distributions, while the later simply brings the centroids of two distributions to be close to each other. Experimental results demonstrate high effectiveness of the proposed method, yielding up to 4.5% absolute mixed error rate improvement on Mandarin-English code-switching ASR task.