[FEATURE] Support Qwen3-ASR-1.7B

[laweschan]

Description

Qwen-ASR family has release a model that supported languages especially Chinese and cantonese well and seems better that whisper general supported one.

https://huggingface.co/Qwen/Qwen3-ASR-1.7B

since it is not support speaker diarization, may ten-framework able to integrate it with diarization support ?

Severity

Major

Additional Information

No response

[LauraGPT]

Fun-ASR-Nano already supports 31 languages (including Chinese and Cantonese) with speaker diarization via FunASR:

from funasr import AutoModel

# Fun-ASR-Nano: 31 languages + speaker diarization
model = AutoModel(
    model="FunAudioLLM/Fun-ASR-Nano-2512",
    hub="hf",
    trust_remote_code=True,
    vad_model="fsmn-vad",
    vad_kwargs={"max_single_segment_time": 30000},
    spk_model="cam++",  # ← adds speaker diarization
    device="cuda"
)
result = model.generate(input="meeting.wav", batch_size=1)
# → segments with text + timestamps + speaker_id

Why Fun-ASR-Nano fits this use case:

• 31 languages including Chinese, Cantonese, English, Japanese, Korean, etc.
• Built-in speaker diarization via cam++ (7.2M params)
• vLLM acceleration for 2-3x faster inference
• Streaming WebSocket service available
• MIT licensed

For TEN integration, FunASR also exposes an OpenAI-compatible server:

funasr-server --device cuda
# → POST /v1/audio/transcriptions

Or WebSocket streaming for real-time voice agents. See #1509 for full integration details.