Qwen3-ASR 1.7B · OpenASR
State-of-the-art multilingual speech recognition across 52 languages & dialects
Native speech-to-text in the OpenASR runtime — engineered for peak performance on CPU & GPU, no Python at inference time.
✨ Highlights
- 🏆 SOTA among open-source ASR — competitive with the strongest proprietary commercial APIs
- 🌍 52 languages & dialects — 30 languages plus 22 Chinese dialects, with spoken-language identification
- 🎧 Robust on hard audio — clean speech, singing voice, and songs over background music
- 🦀 Native in OpenASR —
.oasrpacks run with no Python at inference, engineered for peak performance on CPU & GPU
🚀 Quickstart
# 1. Install the OpenASR CLI · https://openasr.org
# 2. Pull a build (pick a quant — see the table below)
openasr pull qwen3-asr-1.7b:q8
# 3. Transcribe
openasr transcribe audio.wav --model qwen3-asr-1.7b
All builds for this model:
openasr pull qwen3-asr-1.7b:fp16
openasr pull qwen3-asr-1.7b:q8
openasr pull qwen3-asr-1.7b:q4
📦 Available builds
| Quant | File (.oasr) | Size | RAM peak | RTF · M1 CPU | RTF · M1 GPU | JFK ΔWER vs fp16 |
|---|---|---|---|---|---|---|
| fp16 | qwen3-asr-1.7b-fp16.oasr | 4.70 GB | 4.43 GB | 0.90× | 1.63× | 0.0% |
| q8_0 | qwen3-asr-1.7b-q8_0.oasr | 2.51 GB | 4.14 GB | 0.35× | 0.60× | 0.0% |
| q4_k | qwen3-asr-1.7b-q4_k.oasr | 1.33 GB | 4.04 GB | 0.31× | 0.40× | 0.0% |
RTF = real-time factor on the fixed 11s JFK clip (lower is faster); RAM peak measured per pack in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy. q8_0 is the recommended default — near-reference quality at a fraction of the footprint.
🧠 About Qwen3-ASR 1.7B
Qwen3-ASR-1.7B is a multilingual automatic speech recognition model (~2B parameters,
~1.7B active, BF16) from Alibaba's Qwen3-ASR family that transcribes speech while
identifying the spoken language across 30 languages, 22 Chinese dialects, and a range
of regional English accents — and it holds up on hard audio including singing voice and
songs over background music. A single unified checkpoint serves both offline and
real-time streaming transcription, with word-level timestamps available via the
companion Qwen3-ForcedAligner-0.6B; the Qwen team reports state-of-the-art quality among
open-source ASR models and accuracy competitive with commercial APIs. This OpenASR repo
repackages the original Qwen/Qwen3-ASR-1.7B weights as .oasr packs that run natively in
the OpenASR runtime with no Python at inference time. For most users the q8_0 build is the
recommended default — near-reference accuracy at roughly half the footprint — while q4_k
suits tight-memory deployments and fp16 is reserved for verification or maximum fidelity.
⚙️ How these packs were made
Converted from Qwen/Qwen3-ASR-1.7B with the OpenASR importer:
openasr model-pack import qwen <src> <out>.oasr \
--package-id qwen3-asr-1.7b --quantization {fp16,q8-0,q4-k}
The .oasr container is GGUF-backed; packs use zero-copy mmap weight binding and graph
buffer reuse to keep peak memory low.
⚖️ License
These packs inherit the upstream model's license: Apache-2.0 (source). OpenASR packaging retains the upstream copyright and NOTICE; the only modifications are format conversion and quantization.
🙏 Acknowledgements
This pack is a redistribution of Qwen3-ASR-1.7B, created and open-sourced by the Qwen team at Alibaba (Qwen/Qwen3-ASR-1.7B). All credit for the original architecture, training, and weights belongs to them; the license is inherited from and identical to the upstream model (Apache-2.0). Word-level timestamps additionally require the companion Qwen3-ForcedAligner-0.6B. The GGUF quantization recipe and bit-identity verification methodology were informed by cstr/qwen3-asr-1.7b-GGUF. Thank you to both teams for releasing their work openly.
🔗 Links
- 🦀 OpenASR — https://github.com/QuintinShaw/openasr
- 🌐 Website — https://openasr.org
- 🤗 Upstream model — Qwen/Qwen3-ASR-1.7B