qwen3-asr-1.7b-onnx

Qwen3-ASR-1.7B — ONNX (speech-llm)

ONNX export of Qwen/Qwen3-ASR-1.7B-hf for onnx-asr with the speech-llm model type (requires the speech-llm model-family patches, see TigreGotico/onnx-asr PR #3).

Three-graph layout: encoder.onnx (audio -> projected audio embeddings), embed_tokens.onnx, and a merged prefill+decode decoder.onnx with a 28-layer KV cache (hidden size 2048, 8 KV heads, head dim 128). Prompt token ids are baked into config.json, so no tokenizer is needed at runtime. fp32 output matches native transformers character-for-character on FLEURS test clips (en, pt). fp32 and int8 graphs included.

Known limitation (int8)

The int8 decoder can fail on longer utterances: on one 16 s test clip it produced an empty transcript where fp32 was exact (the 0.6B sibling only shows mild drift). Prefer fp32 for audio longer than ~15 s until this is resolved.

Usage

import onnx_asr
model = onnx_asr.load_model("speech-llm", "path/to/this/repo")  # or quantization="int8"
print(model.recognize("audio_16khz.wav"))

Notes

• int8 quantization can drift on longer audio; validate int8 output against fp32 for your use case before relying on it in production. fp32 is character-for-character faithful to the native model on the FLEURS validation clips used for verification.

License: Apache-2.0 (inherited from the source model).