Qwen3 ASR Enhanced

Overview

Qwen3 ASR Enhanced is a fine-tuned variant of Qwen/Qwen3-ASR-1.7B with added support for nonverbal vocalization recognition. In addition to standard speech transcription, this model captures and labels nonverbal sounds such as laughter, sighs, coughs, breaths, throat clearing, and crying inline within the transcript.

This makes the model especially useful for use cases where paralinguistic cues matter: podcast and interview transcription, dialogue analysis, conversational AI training data, accessibility captioning, emotional/affective speech research, and content moderation pipelines where tone and reaction context are as important as the words themselves.

The model retains the strong multilingual ASR performance of the base Qwen3-ASR-1.7B across English, German, French, etc. along with its robustness to noisy environments, accents, and long-form audio.

NOTE: This is an alpha release of the model, a new version will be released soon with additional features such as support for speaker diarization and improved nonverbal tag stability.

Use Cases

• Rich transcription for podcasts, interviews, and meetings where laughter and sighs add meaning
• Dataset creation for training conversational or expressive TTS systems
• Accessibility - generating more descriptive captions for media
• Affective computing & dialogue research where nonverbal cues are signal, not noise
• Content analysis for sentiment, engagement, or moderation workflows

Quickstart

Install the qwen-asr package:

pip install -U qwen-asr

Then load the model and run transcription:

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "mrfakename/Qwen3-ASR-Enhanced-v0.1",
    dtype=torch.bfloat16,
    device_map="cuda:0",
    max_inference_batch_size=32,
    max_new_tokens=256,
)

results = model.transcribe(
    audio="path/to/your/audio.wav",
    language=None,  # auto-detect, or set "English" / "German" / "French"
)

print(results[0].language)
print(results[0].text)
# Example output: "Yeah I mean [laughs] that was honestly the best part [sigh] of the whole thing."

For batch inference, vLLM-backed serving, streaming, or word-level timestamps via the forced aligner, follow the same patterns documented in the base Qwen3-ASR README - this model is a drop-in replacement.

Citation

If you use this model, please also cite the original Qwen3-ASR work:

@article{Qwen3-ASR,
  title={Qwen3-ASR Technical Report},
  author={Xian Shi and others},
  journal={arXiv preprint arXiv:2601.21337},
  year={2026}
}