Fun-ASR: An End-to-End Large Model Framework for Speech Recognition Supporting 31 Languages

Fun-ASR is an open-source end-to-end speech recognition toolkit from Alibaba Tongyi Laboratory. It supports 31 languages, dialect recognition, lyrics recognition, custom hotwords, timestamp generation, speaker diarization, and other functions. Trained on tens of millions of hours of speech data, it represents a significant technological advancement in the current speech recognition field and has notable practical value in global application scenarios.

• Original Author/Maintainer: FunAudioLLM (Speech Team of Alibaba Tongyi Laboratory)
• Source Platform: GitHub
• Release Date: May 25, 2026
• Original Link: https://github.com/FunAudioLLM/Fun-ASR

As an end-to-end speech recognition toolkit, Fun-ASR aims to provide a complete speech recognition solution. Its multilingual support (covering major languages, dialects, and accent variations) makes it suitable for global scenarios.

Fun-ASR covers key needs in the speech recognition field:

1. Multilingual Recognition: Supports 31 languages, dialects, and accent variations;
2. Custom Hotwords: Allows adding professional terms to improve recognition accuracy in specific domains;
3. Timestamp Generation: Provides precise time alignment information, suitable for scenarios like subtitle generation;
4. Speaker Diarization: Distinguishes segments from different speakers in audio, facilitating meeting minutes and more;
5. Lyrics Recognition: Processes speech content in music, applicable to music information retrieval.

Technical Architecture: Based on the deep learning large model paradigm, it uses an end-to-end training approach, directly mapping raw audio to text output. This simplifies system complexity and jointly optimizes all components to improve overall performance.

Training Scale: Trained on tens of millions of hours of speech data, it offers the following advantages:

• Learns robust speech representations to adapt to changes like noise and accents;
• Has zero-shot/few-shot transfer capabilities;
• Provides a foundation for multilingual training, supporting a single model to handle multiple languages.

Fun-ASR has a wide range of application scenarios:

• Content Creation: Video subtitle generation, podcast transcription, real-time live subtitles;
• Intelligent Customer Service: Adapts to international customer service scenarios;
• Education: Speech evaluation, oral practice assistance, classroom recording organization;
• Healthcare: Voice medical record entry;
• Internet of Things: Voice interaction components for smart homes, in-vehicle systems, etc.

For developers, it provides a complete toolchain and pre-trained models, and its open-source nature supports joint improvement by the community.

Summary: Fun-ASR represents an important direction for the evolution of speech recognition into the large model era. Through its end-to-end architecture, large-scale multilingual training, and rich features, it provides a powerful foundational platform for developers and researchers, promoting technological democratization and lowering application barriers.

Outlook: In the future, we will continue to optimize for more language support, lower resource consumption, and higher real-time performance. We look forward to the community developing more innovative voice applications based on this framework.