# FunASR

> FunASR is an open-source AI tool for design and media. Open-source speech recognition toolkit with diarization, for self-hosted and edge. It is available on Linux, macOS, Windows, CLI, API and self-hosted.

- **Kind:** Tool
- **Pricing:** Open source
- **Platforms:** Linux, macOS, Windows, CLI, API, self-hosted
- **Website:** https://www.funasr.com
- **Repository:** https://github.com/modelscope/funasr
- **GitHub stars:** 21,062
- **License:** MIT
- **Categories:** [design and media](https://ai-powered.com/categories/design-and-media), translation quality
- **Listing updated:** 2026-11-18

## About

The README describes FunASR as a Python toolkit that pairs speech recognition with voice activity detection, punctuation, speaker separation and emotion or audio-event models. Its pipelines can be run from a few lines of Python, from a command-line tool, or as an OpenAI-compatible HTTP service that you host yourself.

The project is organised around choosing a task, a checkpoint and a runtime separately. The README's table pairs options such as SenseVoiceSmall, Fun-ASR-Nano, Paraformer streaming and a llama.cpp/GGUF build for CPU and edge devices, and notes that each has different language coverage and speaker behaviour. Speaker labels are anonymous indices within a recording, not identities.

It suits teams that need transcription they can run on their own hardware, including developers building subtitle, meeting-note or voice-agent pipelines. Model weights carry their own licences that are separate from the MIT-licensed toolkit code, so those should be checked per model.

## Features

- **Speech recognition models:** Offers several ASR checkpoints covering Chinese, English, Japanese, Korean and more, with a 31-language option.
- **Speaker-aware segmentation:** Combines voice activity detection with speaker embeddings to label segments with anonymous speaker indices.
- **Streaming transcription:** Supports chunked real-time transcription using a streaming checkpoint and per-session cache.
- **OpenAI-compatible server:** Can be run as a local HTTP service exposing transcription endpoints that follow the OpenAI API shape.
- **CPU and edge builds:** Provides a llama.cpp/GGUF runtime for running supported models without Python on CPU or GPU via Vulkan or CUDA.
- **Command-line interface:** Transcribes audio files from the terminal with options for JSON or SRT output and speaker timestamps.

## Good for

- Generating subtitles for batches of audio or video files
- Transcribing meetings with speaker separation on self-hosted hardware
- Adding speech input to voice agents through an OpenAI-style API
- Running offline transcription on CPU-only or edge devices
- Comparing ASR checkpoints against Whisper or cloud providers on your own recordings

## FAQ

### Is FunASR free to use?

The README states the toolkit source code is released under the MIT License. Pretrained model weights are licensed separately, so each model card should be checked.

### Does it identify speakers by name?

No. The README says speaker indices are local to a recording and are not known-person identities.

### Which Python version is required?

The README lists Python 3.8 or newer for installing from source, and its deployment example uses Python 3.11.

### Does the HTTP server include authentication?

The README says the local service it describes is unauthenticated and should stay on loopback unless the security guide is followed.

### Is FunASR free?

Yes. FunASR is open source, so you can use and self-host it for free.

## Where this comes from

- Description: written from the project's README and the project's website.

---
Source: https://ai-powered.com/tools/funasr-2
