FunASR: Open-Source Speech Recognition Toolkit — Complete Guide
Discover FunASR, the open-source end-to-end speech recognition toolkit by ModelScope. Learn about streaming ASR, VAD, speaker diarization, and how to deploy it with Docker.
FunASR is an open-source, end-to-end speech recognition toolkit developed by ModelScope (Alibaba DAMO Academy). With over 19,500 GitHub stars, it has quickly become one of the most popular frameworks for building, training, and deploying automatic speech recognition (ASR) systems — from streaming real-time transcription to offline batch processing.
🚀 Want to deploy FunASR yourself?
Docker configs, system requirements, and installation guides — all on one page.
View FunASR Tool Page →What Makes FunASR Stand Out?
Unlike many ASR toolkits that focus on a single task, FunASR provides a complete pipeline covering the full speech recognition lifecycle:
| Feature | Description |
|---|---|
| 🎤 Streaming ASR | Real-time speech-to-text with low latency, ideal for live captioning and voice assistants |
| 🔇 Voice Activity Detection (VAD) | Automatically detects when speech starts and stops, filtering silence |
| 🔣 Punctuation Restoration | Adds commas, periods, and question marks to transcribed text |
| 👥 Speaker Diarization | Identifies \"who spoke when\" — essential for meeting transcription |
| 😊 Emotion Recognition | Detects emotional tone from speech (happy, sad, angry, etc.) |
OpenAI-Compatible & MCP APIs
One of FunASR's most compelling features is its OpenAI-compatible API. This means any application built for OpenAI's Whisper API can switch to a self-hosted FunASR instance with minimal changes. It also exposes a Model Context Protocol (MCP) server, letting AI agents call speech recognition directly as a tool.
Architecture Overview
FunASR is built on PyTorch and leverages state-of-the-art models including:
- Paraformer — non-autoregressive end-to-end ASR for fast inference
- FSMN-VAD — efficient voice activity detection
- CT-Transformer — punctuation restoration
- CAM++ / ResNet — speaker embedding for diarization
The toolkit supports both training from scratch and fine-tuning pre-trained models on custom datasets, making it suitable for research and production alike.
Supported Languages
While FunASR has particularly strong support for Chinese (Mandarin and Cantonese), it also supports multilingual ASR including English, Japanese, Korean, and more through its Paraformer-large model.
Use Cases
- Live captioning for meetings, webinars, and streaming
- Voice assistants with real-time speech understanding
- Call center analytics with speaker attribution
- Content generation — transcribe podcasts, lectures, and interviews
- Accessibility tools for hearing-impaired users
🚀 Ready to try FunASR?
Get the Docker setup, system requirements, and deployment guide.
View FunASR Tool Page →