SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Fun-ASR - The speech recognition large model jointly launched by DingTalk and Tongyi

Author: AI小集 · Original source: AI工具集

What is Fun-ASR

Fun-ASR is a new-generationspeech recognitionlarge model jointly launched by DingTalk and the speech team of Tongyi Lab. Trained on massive audio data, it can accurately recognize professional terminology from more than a dozen industries such as internet, technology, home decoration, and animal husbandry, and can understand "jargon." For example, in the insurance industry, accuracy improved by 18% compared with the past, and in industries such as home decoration and animal husbandry, it also achieved improvements of 15%-20%. The model can combine enterprise information within DingTalk for reasoning optimization, reducing hallucination issues and providing more reliable transcription results. Fun-ASR supports customized training of enterprise-specific models, and can use the enterprise's real speech data to further optimize the algorithm, improve the recognition accuracy of dedicated vocabulary, and support the import of up to 1,000+ hotwords.

Currently, Fun-ASR has been integrated into multiple functional modules of DingTalk, including meeting subtitles, smart meeting notes, and voice assistant, providing a stable, efficient, and easily scalable speech recognition solution for enterprise-level contexts.

Tongyi Lab has comprehensively upgraded the core capabilities of Fun-ASR. The recognition accuracy in noisy scenarios has increased to 93%, it supports free mixing of 31 languages, lyrics and rap recognition, and reduces the first-token latency of the streaming recognition model to 160ms, significantly improving speech recognition performance in complex environments. In addition, the lightweight version Fun-ASR-Nano-0.8B has been officially open-sourced, with the total parameter count compressed to 0.8B, lower inference cost, and support for local deployment and customized fine-tuning, providing developers with an efficient and flexible speech recognition solution.

Fun-ASR

Main features of Fun-ASR

  • Multi-industry terminology recognition: Fun-ASR has been trained on massive audio data and can accurately recognize professional terminology from more than a dozen industries such as internet, technology, home decoration, animal husbandry, and automotive. In actual testing, accuracy in the insurance industry improved by 18% compared with the past, and in industries such as home decoration and animal husbandry it improved by 15%-20%. It supports the import of up to 1,000+ hotwords to further optimize recognition of uncommon vocabulary.
  • Context-aware optimization: The model can combine enterprise information within DingTalk (such as contacts, schedules, knowledge base, etc.) for reasoning optimization, effectively alleviating the hallucination issues that large models may have, and providing more reliable transcription results. It takes effect only after enterprise authorization.
  • Enterprise-specific customized training: Based on an efficient end-to-end training architecture, Fun-ASR can optimize the algorithm based on real-scenario speech data provided by the enterprise, improving the recognition accuracy of dedicated vocabulary such as brand names, project codes, product names, and people's names.
  • Multi-scenario integrated applications: Fun-ASR has been integrated into multiple functional modules of DingTalk, including meeting subtitles and simultaneous interpretation, smart meeting notes, and voice assistant, providing a stable, efficient, and easily scalable speech recognition foundation for enterprise-level contexts and meeting enterprises' high requirements for speech recognition.

Technical principles of Fun-ASR

  • Massive data training: Fun-ASR has been trained on hundreds of millions of hours of audio data, covering multiple industries and scenarios, and can accurately understand professional terminology in different fields.
  • Industry co-creation optimization: Through co-creation with DingTalk's multi-industry customers based on real scenarios, the model performs excellently in more than a dozen fields such as internet, technology, home decoration, animal husbandry, and automotive, significantly improving the recognition accuracy of professional terminology.
  • Contextual reasoning optimization: The model can combine the enterprise's existing information within DingTalk (such as contacts, schedules, knowledge base, etc.) for reasoning optimization, effectively alleviating the hallucination issues that large models may produce, and providing more reliable transcription results.
  • End-to-end training architecture: Based on an efficient end-to-end training architecture, Fun-ASR can leverage real-world speech data provided by enterprises to further optimize the algorithm, improve the recognition accuracy of proprietary vocabulary, and support customized training of enterprise-specific models.
  • Custom hotword support: Provides enterprise custom hotword capabilities, supporting the import of up to 1,000+ hotwords, further optimizing the recognition of rare vocabulary and proprietary terminology.

Fun-ASR Project Address

  • GitHub Repository：https://github.com/FunAudioLLM/Fun-ASR
  • HuggingFace Model Library：https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512

Fun-ASR Application Scenarios

  • Meeting subtitles and simultaneous interpretation: Fun-ASR can transcribe meeting content in real time, providing accurate subtitles and simultaneous interpretation services for meetings, helping participants better understand and record key meeting points.
  • Smart meeting minutes: The model can automatically generate meeting minutes, extract key information and action items, saving manual organizing time and improving meeting efficiency.
  • Voice assistant: Supports voice commands and interaction. Users can complete various operations through voice commands, such as querying information, scheduling appointments, etc., enhancing the user experience.
  • Home decoration and animal husbandry industries: In home decoration enterprises such as Kuka Home, the model can accurately recognize professional expressions such as “Belgian imported Pulse latex,” providing a reliable basis for subsequent analysis of customer needs. In the animal husbandry industry, it can also accurately recognize related terminology, helping enterprises operate efficiently.
  • Insurance industry: The application of Fun-ASR in the insurance industry significantly improves the accuracy of speech recognition, helping insurance enterprises better handle customer inquiries and business processes.