Fun-ASR-Nano-API

中文文档

OpenAI-compatible Speech-to-Text API powered by Fun-ASR-Nano (Alibaba FunAudioLLM).

800M parameters. 31 languages. Chinese dialects. Hotwords. VAD. Punctuation. Speaker diarization. One container.

Features

• OpenAI-compatible /v1/audio/transcriptions endpoint
• 31 languages (Chinese, English, Japanese, Korean, Vietnamese, Arabic, and more)
• Chinese dialects: Wu, Cantonese, Min, Hakka, Gan, Xiang, Jin + 26 regional accents
• Hotword boosting (improve recognition of domain-specific terms)
• VAD (Voice Activity Detection) — auto-segments long audio
• Automatic punctuation restoration
• Speaker diarization (who said what)
• Switchable models via MODEL_ID:
  • Fun-ASR-Nano-2512 (default) — full features, dialects, hotwords
  • Fun-ASR-MLT-Nano-2512 — 31 languages including European

Quick Start

# Default: Fun-ASR-Nano (Chinese dialects, hotwords, full features)
docker run -d --gpus all \
  -p 8080:8080 \
  -v /mnt/user/appdata/fun-asr-nano-api/models:/root/.cache/huggingface \
  -e MODEL_ID=FunAudioLLM/Fun-ASR-Nano-2512 \
  --shm-size=4g \
  --name fun-asr-nano-api \
  ghcr.io/hsiang-han/fun-asr-nano-api:latest

# With speaker diarization enabled
docker run -d --gpus all \
  -p 8080:8080 \
  -v /mnt/user/appdata/fun-asr-nano-api/models:/root/.cache/huggingface \
  -e MODEL_ID=FunAudioLLM/Fun-ASR-Nano-2512 \
  -e ENABLE_SPK=true \
  --shm-size=4g \
  --name fun-asr-nano-api \
  ghcr.io/hsiang-han/fun-asr-nano-api:latest

# Alternative: Fun-ASR-MLT (31 languages including European)
docker run -d --gpus all \
  -p 8080:8080 \
  -v /mnt/user/appdata/fun-asr-nano-api/models:/root/.cache/huggingface \
  -e MODEL_ID=FunAudioLLM/Fun-ASR-MLT-Nano-2512 \
  --shm-size=4g \
  --name fun-asr-nano-api \
  ghcr.io/hsiang-han/fun-asr-nano-api:latest

China users: add -e HF_ENDPOINT=https://hf-mirror.com.

Usage Examples

# Basic transcription (OpenAI-compatible)
curl -X POST http://localhost:8080/v1/audio/transcriptions \
  -F "[email protected]"

# Specify language
curl -X POST http://localhost:8080/v1/audio/transcriptions \
  -F "[email protected]" \
  -F "language=中文"

# With hotwords (boost domain-specific terms)
curl -X POST http://localhost:8080/v1/audio/transcriptions \
  -F "[email protected]" \
  -F "hotwords=人工智能,大语言模型,通义千问"

# Verbose output (timestamps + speaker info if enabled)
curl -X POST http://localhost:8080/v1/audio/transcriptions \
  -F "[email protected]" \
  -F "response_format=verbose_json"

Response Examples

Standard (json):

{"text": "今天天气真好，我们出去玩吧。"}

Verbose (verbose_json, with speaker diarization):

{
  "text": "今天天气真好，我们出去玩吧。",
  "language": "中文",
  "segments": [
    {"start": 0.0, "end": 2.5, "text": "今天天气真好，", "speaker": 0},
    {"start": 2.5, "end": 4.8, "text": "我们出去玩吧。", "speaker": 1}
  ]
}

API Endpoints

Endpoint	Method	Description
/v1/audio/transcriptions	POST	Speech-to-text (OpenAI-compatible)
/v1/models	GET	List models
/health	GET	Health check (shows enabled features)
/docs	GET	Swagger documentation

Environment Variables

Variable	Default	Description
MODEL_ID	FunAudioLLM/Fun-ASR-Nano-2512	Model to load
DEVICE	cuda:0	Compute device (cuda:0, cpu)
LANGUAGE	auto	Default language (auto, 中文, English, 日文, etc.)
ENABLE_VAD	true	Voice Activity Detection (segments long audio)
ENABLE_PUNC	true	Auto punctuation restoration
ENABLE_SPK	false	Speaker diarization (who said what)
PORT	8080	API server port
HF_ENDPOINT	https://huggingface.co	HuggingFace mirror

Available Models

Model ID	Languages	Features
FunAudioLLM/Fun-ASR-Nano-2512	Chinese+dialects, English, Japanese, Korean, + more	Hotwords, dialects, accents, lyrics
FunAudioLLM/Fun-ASR-MLT-Nano-2512	31 languages (including European)	Broader language coverage

Pipeline Components

Component	Model	Size	Enabled by
ASR	Fun-ASR-Nano-2512	800M	always
VAD	fsmn-vad	0.4M	ENABLE_VAD=true
Punctuation	ct-punc	290M	ENABLE_PUNC=true
Speaker	cam++	7.2M	ENABLE_SPK=true

Hardware Requirements

• NVIDIA GPU with 2GB+ VRAM (ASR only) or 3GB+ (with all components)
• NVIDIA driver 550+
• Docker with NVIDIA Container Toolkit

Credits

• Fun-ASR-Nano by Alibaba FunAudioLLM / Tongyi Lab
• FunASR toolkit

License

Apache-2.0