Speech recognition for 52 languages, plus a forced-alignment sidekick

What it does

Qwen3-ASR is a family of open-source speech recognition models from Alibaba’s Qwen team. The two main variants—1.7B and 0.6B parameters—transcribe and identify languages across 52 languages and Chinese dialects, including audio with singing or background music. A separate 0.6B non-autoregressive forced-alignment model predicts timestamps for text–speech pairs in 11 languages.

The interesting bit

The 1.7B model reportedly achieves state-of-the-art results among open-source ASR models and is competitive with proprietary commercial APIs. The 0.6B variant is tuned for throughput, with the project claiming it reaches 2000× throughput at a concurrency of 128—though the baseline for that multiplier is unspecified. The forced aligner is a distinct, non-autoregressive model rather than an afterthought bolted onto the main ASR.

Key highlights

• Supports 30 languages plus 22 Chinese dialects and English regional accents
• Handles speech, singing, and songs with background music in a single model
• Unified streaming and offline inference without swapping models
• Ships with an inference toolkit supporting vLLM batching, async serving, and timestamp prediction
• Available as a Python package with both transformers and vLLM backends, plus an official Docker image

Caveats

• FlashAttention 2 acceleration requires compatible hardware and only works with float16 or bfloat16 weights
• The vLLM backend can trigger Python multiprocessing spawn errors without proper entry-point guards

Verdict

Developers building multilingual voice apps or subtitle pipelines should look here, especially if they need a single model that handles both Mandarin dialects and music. If you only need English-only transcription on CPU, this is overkill.