qwen-asr

What it is and what it does

Qwen3-ASR is a Python package wrapping Qwen's speech recognition models (1.7B and 0.6B variants) that transcribe audio to text in 52 languages and dialects, identify the spoken language automatically, and optionally align text to speech timestamps. It supports both streaming and offline inference, handles various audio formats (speech, singing, songs with background music), and provides multiple backends: a transformers-based inference engine and an optional vLLM backend for faster batch processing.

The package is designed for developers who need multilingual ASR in production or research settings. It depends on transformers, librosa, soundfile, sox, and other audio/ML libraries; optional vLLM support requires additional installation. The forced-aligner model (Qwen3-ForcedAligner-0.6B) is a separate, non-autoregressive component for timestamping arbitrary text units within speech up to 5 minutes long across 11 languages.

Use it for:

• Transcribe multilingual customer support calls or meeting recordings without language detection overhead
• Build a voice-to-text pipeline for content in Chinese dialects, Arabic, or other underrepresented languages
• Generate precise word-level timestamps for subtitle generation or speech-to-text alignment in video editing
• Deploy a real-time streaming transcription service using the vLLM backend for high-concurrency scenarios
• Fine-tune the model on domain-specific audio (medical, legal, technical) using the provided training framework

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Qwen3-ASR provides speech recognition and language identification for 52 languages and dialects, plus forced-alignment for timestamping speech in 11 languages, with both streaming and offline inference modes.

Yes, with conditions. Install if you need multilingual ASR (especially for Chinese dialects or non-English languages) and can tolerate a fresh Python 3.9+ environment with 11 runtime dependencies. The package is young (released 2026-01-29) and marked aging, so expect possible API changes; no known security vulnerabilities. Permissive Apache-2.0 license poses no legal barrier. Requires GPU for practical inference speed and manual model download in offline environments.

Install

qwen-asr on PyPI

pip install qwen-asr

uv add qwen-asr

poetry add qwen-asr

Installing qwen-asr

Before you install

Low friction: pure Python wheel with no compiled dependencies. Aging status (196 days since release) but recent activity; supports Python 3.9–3.13. Eleven runtime dependencies include transformers, librosa, soundfile, sox, and optional vLLM backend—typical for ML inference packages.

License in practice

Apache-2.0 permissive license allows commercial and private use with minimal restrictions; attribution required but no copyleft obligations.

Quickstart

pip install qwen-asr
from qwen_asr import Qwen3ASRModel
model = Qwen3ASRModel.from_pretrained('Qwen/Qwen3-ASR-0.6B')
result = model.transcribe('audio.wav')

Requires Python ≥3.9. Model weights download automatically; offline environments need manual download via modelscope or huggingface-cli. GPU recommended for reasonable inference speed.

Verify before relying

• Actual throughput and latency figures for the 0.6B model at stated concurrency levels
• Accuracy benchmarks against proprietary APIs mentioned in description
• Memory requirements for each model variant on typical hardware
• Whether vLLM backend is required for streaming inference or optional

Package facts

Evidence: qwen_asr-0.0.6-py3-none-any.whl

Tags

More Artificial Intelligence packages