FunASR Models

Choose the right model for your use case — from ultra-fast multilingual recognition to the highest Chinese accuracy.

Quick Comparison

Model	Speed	Languages	Params	Best For
Fun-ASR-Nano ⭐	vLLM; benchmark your hardware	zh/en/ja + Chinese dialects/accents	800M	Flagship · LLM-ASR · hardest cases
Fun-ASR-MLT-Nano	vLLM	31	800M	Separate multilingual checkpoint
SenseVoice Small	Non-autoregressive; benchmark your hardware	zh, yue, en, ja, ko	234M	Five-language ASR, emotion detection
Paraformer-zh	Non-autoregressive; benchmark your hardware	zh, yue	220M	Best Chinese accuracy
MOSS-Transcribe-Diarize	Batch; NVIDIA GPU	See official model card	Third-party	Long-form transcription + speaker labels
cam++	realtime	any	7.2M	Speaker diarization & verification

ASR Models

Fun-ASR-Nano ⭐ Flagship / default

800M params · LLM-based (SenseVoice encoder + Qwen3-0.6B) · GitHub · HuggingFace

Next-generation LLM-based ASR model. Combines SenseVoice's audio encoder with Qwen3-0.6B language model for superior context understanding. Supports vLLM acceleration for high-throughput batch inference and real-time streaming. The released model.pt checkpoint does not provide reliable checkpoint-native character timestamps (issue #106).

vLLM accelerated zh/en/ja + Chinese dialects/accents Streaming LLM-quality

When to use

Best for high-throughput batch processing, real-time subtitles, and scenarios where LLM-quality context understanding improves output (e.g., proper nouns and code-switching). For reliable character-level timestamps, use Paraformer.

# With vLLM acceleration
from funasr import AutoModel
model = AutoModel(model="FunAudioLLM/Fun-ASR-Nano-2512", device="cuda", backend="vllm")
result = model.generate(input="audio.wav")
print(result[0]["text"])

Fun-ASR-MLT-Nano 31 languages

800M params · Separate multilingual checkpoint · HuggingFace · ModelScope

A separate checkpoint for broad multilingual recognition across 31 languages. Its model ID and language scope differ from flagship Fun-ASR-Nano.

vLLM accelerated 31 languages Separate checkpoint

When to use

Choose MLT-Nano for recognition across 31 languages; choose flagship Nano for zh/en/ja and Chinese dialects/accents.

from funasr import AutoModel
model = AutoModel(model="FunAudioLLM/Fun-ASR-MLT-Nano-2512", device="cuda")
result = model.generate(input="audio.wav")

SenseVoice Small

234M params · Non-autoregressive · GitHub · HuggingFace

Ultra-fast speech recognition with built-in emotion and audio event detection. Supports Mandarin, Cantonese, English, Japanese, and Korean. Its non-autoregressive architecture suits low-latency deployment; throughput depends on hardware, audio length, and batch configuration.

Non-autoregressive Five languages Emotion detection Audio events CPU-viable

When to use

Best for: multilingual applications, real-time streaming, batch processing large audio collections, applications needing emotion or audio event detection.

from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall")
result = model.generate(input="audio.wav")
print(result[0]["text"])

Paraformer-zh Large

220M params · Non-autoregressive · HuggingFace

Highest-accuracy Chinese speech recognition model. Non-autoregressive with CTC-guided attention, trained on 60,000+ hours of Mandarin speech. Includes built-in punctuation restoration and timestamp prediction.

Non-autoregressive; benchmark your hardware Chinese + Cantonese Best accuracy Timestamps Punctuation

When to use

Best for: Chinese-only applications where accuracy is the top priority — meeting transcription, subtitle generation, voice input, training data annotation.

from funasr import AutoModel
model = AutoModel(
    model="iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
    vad_model="iic/speech_fsmn_vad_zh-cn-16k-common-pytorch",
    punc_model="iic/punc_ct-transformer_cn-en-common-vocab471067-large",
)
result = model.generate(input="audio.wav")
print(result[0]["text"])

MOSS-Transcribe-Diarize third-party model

Published by OpenMOSS · Model ID: OpenMOSS-Team/MOSS-Transcribe-Diarize · HuggingFace · FunASR deployment guide

A third-party Apache-2.0 model published by OpenMOSS that produces long-form transcription, timestamps, and speaker labels in one offline request. FunASR maps structured results for AutoModel, vLLM, and SGLang Omni. It does not require an external VAD service, which does not mean the model has no internal segmentation.

Transcription + speaker labels No external VAD vLLM / SGLang Omni Non-realtime

When to use

Long-form offline transcription for multi-speaker meetings, interviews, podcasts, and calls. This is not FunASR-owned model weight; validate overlap, timestamps, speaker consistency, and capacity on real multi-speaker audio before production.

Supporting Models

cam++ (Speaker Diarization)

7.2M params · HuggingFace

Lightweight speaker embedding model for speaker diarization (who spoke when) and speaker verification (is this the same person). Only 7.2M parameters — runs on CPU in realtime.

7.2M params Diarization Verification CPU realtime

FSMN-VAD

Built-in · Voice Activity Detection

Feedforward Sequential Memory Network for Voice Activity Detection. Accurately detects speech segments in audio, handling silence, noise, and music. Used as a preprocessing step for all ASR models.

VAD Lightweight

CT-Transformer (Punctuation)

Built-in · Punctuation Restoration

Automatically adds punctuation to ASR output — commas, periods, question marks, etc. Supports Chinese and English. Dramatically improves readability of transcription output.

Punctuation zh + en

OpenAI-Compatible API

All models are available through funasr-server, which exposes an OpenAI-compatible /v1/audio/transcriptions endpoint:

# Start the server
pip install funasr vllm fastapi uvicorn python-multipart
funasr-server --device cuda --port 8000

# Use with any OpenAI-compatible client
curl http://localhost:8000/v1/audio/transcriptions \
  -F file=@audio.wav \
  -F model=SenseVoiceSmall

Drop-in replacement: Any application using OpenAI's Whisper API can switch to FunASR by changing the base URL. No code changes needed — same API format, same response structure.

Deployment Options

Method	Command	Best For
pip	`pip install funasr && funasr-server`	Development, quick testing
Docker	`docker run -d --gpus all -p 8000:8000 ...`	Production deployment
Python API	`from funasr import AutoModel`	Embedding in applications
ONNX	Via Sherpa-ONNX	Mobile, edge, browser