You can deploy this model and start earning money today!

thieunv/
sherpa-onnx-qwen3-asr-1.7B-int8

Name: sherpa-onnx-qwen3-asr-1.7B-int8
Availability: InStock

sherpa-onnx

onnx

automatic-speech-recognition

Views

No views yet

sherpa-onnx-qwen3-asr-1.7B-int8

INT8 ONNX build of Qwen3-ASR-1.7B in the sherpa-onnx from_qwen3_asr layout, for on-device multilingual speech recognition (no Python at runtime).

Base model: Qwen/Qwen3-ASR-1.7B (Apache-2.0)
Exported to sherpa-onnx format with Wasser1462/Qwen3-ASR-onnx (export_qwen3_asr_onnx.py --device cpu --max-total-len 512, INT8 dynamic quantization).
Languages: 30+ incl. Chinese (+dialects), English, Japanese, Vietnamese, Korean, Cantonese, …
Higher accuracy than the 0.6B build at the cost of size/speed.

Files

conv_frontend.onnx        # log-mel / conv frontend
encoder.int8.onnx         # audio encoder (int8)
decoder.int8.onnx         # text decoder w/ KV cache (int8, ~2 GB)
tokenizer/vocab.json
tokenizer/merges.txt
tokenizer/tokenizer_config.json

Usage (sherpa-onnx)

python

import sherpa_onnx
rec = sherpa_onnx.OfflineRecognizer.from_qwen3_asr(
    conv_frontend="conv_frontend.onnx",
    encoder="encoder.int8.onnx",
    decoder="decoder.int8.onnx",
    tokenizer="tokenizer",
    num_threads=4,
)
s = rec.create_stream()
s.accept_waveform(16000, samples_f32_mono)  # 16 kHz mono float32
rec.decode_stream(s)
print(s.result.text)

Requires sherpa-onnx >= 1.13.4 (adds from_qwen3_asr).

License

Apache-2.0, inherited from Qwen/Qwen3-ASR-1.7B. Attribution to the Qwen team (Alibaba Cloud) and the sherpa-onnx / export-script authors.

thieunv/sherpa-onnx-qwen3-asr-1.7B-int8

sherpa-onnx-qwen3-asr-1.7B-int8

Files

Usage (sherpa-onnx)

License

thieunv/
sherpa-onnx-qwen3-asr-1.7B-int8