#Fun-ASR Series

Notes

• Model background: End-to-end speech recognition foundation model released by Tongyi Lab. Pre-trained on tens of millions of hours of real speech data, featuring strong contextual understanding and domain adaptability
• Features: All models are non-streaming, support punctuation, support timestamps. Support low-latency real-time transcription, with recognition accuracy reaching 93% in far-field, high-noise environments
• Version identifier meanings:
  • int8: INT8 quantized version, smaller size, faster inference, suitable for edge deployment
  • LLM: Large model enhanced version, stronger context understanding, suppresses recognition hallucinations
  • CTC: Lightweight classic CTC architecture version, lightweight inference
  • MLT: Multilingual general-purpose version, covers 31 languages
  • split-adaptor: Version with feature adaptation module deployed separately
• Language and capability notes:
  • Fun-ASR-Nano: Supports Chinese, English, Japanese; 7 dialects (Wu, Cantonese, Min, Hakka, Gan, Xiang, Jin); 26 regional accents (Henan, Shanxi, Hubei, Sichuan, Chongqing, Yunnan, Guizhou, Guangdong, Guangxi, Shaanxi, Hebei, Shandong, Anhui, Tianjin, Ningxia, Liaoning, Gansu, Hunan, Heilongjiang, Jilin, Inner Mongolia, Jiangsu, Zhejiang, Fujian, Jiangxi, Hainan); additionally supports lyrics recognition and rap speech recognition
  • Fun-ASR-MLT-Nano: Supports 31 languages total: Chinese, English, Cantonese, Japanese, Korean, Vietnamese, Indonesian, Thai, Malay, Filipino, Arabic, Hindi, Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, Greek, Hungarian, Irish, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Swedish
• Domain advantages: Excellent performance in vertical fields such as education and finance, accurately recognizes domain-specific terminology, effectively suppresses hallucinations and language confusion

#Fun-ASR-Nano Models

#Fun-ASR-MLT-Nano Models