Which FunASR Model? Nano vs MLT-Nano vs SenseVoice vs Paraformer (2026 Guide)

Updated 2026-07-15 · FunASR Team

FunASR ships three main model families, plus a separate multilingual MLT-Nano checkpoint. In one line: with a GPU, use flagship Fun-ASR-Nano for Chinese, English, Japanese, and Chinese dialects/accents; use Fun-ASR-MLT-Nano when you need 31 languages; on CPU use SenseVoice, or Paraformer for Chinese timestamps/hotwords.

Pick at a glance

Your situation	Pick	Why
GPU / zh, en, ja, or Chinese dialects and accents / want the best	⭐ Fun-ASR-Nano	LLM-ASR flagship, strongest on context and hard cases, Chinese CER 8.06%
Need broad multilingual coverage	Fun-ASR-MLT-Nano	Separate checkpoint covering 31 languages
CPU / edge / want fastest	SenseVoice	non-autoregressive & very fast, real-time on CPU, multilingual + emotion, CER 7.81%
CPU + Chinese-only + timestamps/hotwords	Paraformer	lightest, character timestamps, hotwords, CER 10.18%

The Chinese CER figures compare Nano, SenseVoice, and Paraformer; all three beat Whisper on Chinese (~20%) by a wide margin. SenseVoice has a slightly lower CER, but Fun-ASR-Nano is the LLM-based flagship — more robust on real-world audio, context, and proper nouns. MLT-Nano is the choice for 31-language coverage; do not apply its language claim to Nano.

Fun-ASR-Nano — LLM-ASR flagship, the default (GPU)

SenseVoice encoder + a Qwen3-0.6B decoder, supporting Chinese, English, Japanese, 7 Chinese dialect groups, and 26 regional accents, with LLM-grade context understanding for hard cases and long-tail vocabulary.

pip install funasr

from funasr import AutoModel

model = AutoModel(model="FunAudioLLM/Fun-ASR-Nano-2512", disable_update=True, device="cuda")
res = model.generate(input="audio.wav")

print(res[0]["text"])
# 欢迎大家来体验达摩院推出的语音识别模型。  ("Welcome everyone to try the speech model from DAMO Academy")

Models: Hugging Face · ModelScope. For scale/throughput, use vLLM acceleration (~340× real-time measured). See the Fun-ASR-Nano guide.

Fun-ASR-MLT-Nano — separate 31-language checkpoint (GPU)

Choose MLT-Nano for recognition across 31 languages. It is a separate checkpoint from flagship Nano, with a different model ID and language scope.

Models: Hugging Face · ModelScope.

SenseVoice — the CPU/edge choice

Non-autoregressive, lightweight and very fast — real-time even on CPU — emitting language + emotion + audio events in one pass; q8 quantization is ~250MB. Pick it for CPU deployment or lowest latency.

from funasr import AutoModel
from funasr.utils.postprocess_utils import rich_transcription_postprocess

model = AutoModel(model="iic/SenseVoiceSmall", disable_update=True)  # CPU by default
res = model.generate(input="audio.wav", language="auto", use_itn=True)

print(rich_transcription_postprocess(res[0]["text"]))
# 欢迎大家来体验达摩院推出的语音识别模型。

Paraformer — CPU + Chinese-only (timestamps/hotwords)

The lightest Chinese recognizer, with character-level timestamps and hotword customization — use it for subtitles or word-level alignment.

from funasr import AutoModel

model = AutoModel(model="paraformer-zh", vad_model="fsmn-vad", punc_model="ct-punc")
res = model.generate(input="audio.wav", batch_size_s=300)

print(res[0]["text"])
# 欢迎大家来体验达摩院推出的语音识别模型。
# res[0]["timestamp"] -> [[880, 1120], [1120, 1360], ...] per-character start/end ms

Quick decision

Need 31-language coverage? → Fun-ASR-MLT-Nano.
Have a GPU and mainly need zh/en/ja or Chinese dialects/accents? → Fun-ASR-Nano (the default flagship).
CPU only? Chinese-only with timestamps/hotwords → Paraformer; otherwise (incl. multilingual/emotion) → SenseVoice.
Need Cantonese? SenseVoice supports it natively — see Cantonese speech recognition.

More: Chinese speech recognition · FunASR vs Whisper benchmark · Fun-ASR-Nano guide.

The whole FunASR stack is open-source (MIT) — Fun-ASR-Nano / MLT-Nano + SenseVoice + Paraformer + VAD/punctuation/speaker. If it helps, a GitHub Star supports the project 👇

⭐ Star Fun-ASR

Also star:FunASR · SenseVoice · FunClip