中文说明 · Deployment · Reproduction · Validation
FP16 QNN context binaries and a standalone Android runner converted from Qwen/Qwen3-ASR-0.6B-hf. The tested target is Snapdragon 8 Gen 3 / SM8650 / Hexagon HTP V75, using QAIRT 2.51.40.260930 on Samsung SM-F741U with Android 14.
Download the runtime files into one directory and run qwen_runner with the matching Qualcomm runtime. This repository uses a native QNN runtime; the context binaries are loaded by the included runner rather than Transformers from_pretrained.
| Property | Value |
|---|---|
| Source checkpoint | Qwen/Qwen3-ASR-0.6B-hf |
| Source revision | 7f1569a48a89f3e3f4dc3a5c9d28bddd903bc76c |
| Conversion | Fixed-shape FP16 QNN contexts; no additional fine-tuning |
| Target | SM8650, HTP V75, Android AArch64 |
| SDK used | QAIRT 2.51.40.260930 |
| Native build | Android NDK r26c, API 31, C++17 |
| Decoder | 28 layers in 7 blocks; prefill length 16, decode length 1 |
| KV capacity | 1024 positions, including prompt/audio and generated tokens |
| Input | 16 kHz mono little-endian float32 PCM |
| Explicit language prompts | auto, zh, en, ja, pt |
The repository keeps the upstream model's 0.6B product name. Parameter totals displayed by hosting sites may include additional components and differ from that name.
audio_cnn.bin, audio_window.bin: audio encoder contexts.decoder_0_s1.bin … decoder_6_s1.bin: single-token decoder contexts.decoder_0_s16.bin … decoder_6_s16.bin: prefill decoder contexts.output_head.bin: vocabulary projection context.embedding.fp16: embedding table.mel_filters.f32, frontend.json, token_bytes.bin, token_offsets.i32, prompt_*.i32: CPU frontend and tokenizer assets.qwen_runner: standalone Android AArch64 executable; manifest.json records runtime file sizes and provenance.config.json: deployment configuration describing the QNN format, target, runtime, audio input and decoder shapes.source.zip: native C++ engine, export/compile/package scripts, locked Python dependencies and numerical alignment summaries.The package contains 17 contexts. The two mel-check-* synthetic diagnostic arrays from development are omitted from the published runtime manifest. Qualcomm SDK/runtime libraries are installed separately under their own terms. All model assets remain together in the model directory.
Download the full repository, including config.json, into a local model/ directory:
hf download luyii/Qwen3-ASR-0.6B-QNN-FP16-SM8650 --local-dir model
For manual downloads, include config.json together with the runtime files listed in manifest.json. Hugging Face's default download statistics count GET/HEAD requests to configuration query files such as config.json; these are request counts, not confirmed full-package downloads or unique users. See Hugging Face download statistics.
See DEPLOYMENT.md for SDK library preparation and ADB commands. Inside an Android shell:
cd /data/local/tmp/qwen3-asr
export LD_LIBRARY_PATH="$PWD/runtime"
export ADSP_LIBRARY_PATH="$PWD/runtime;/vendor/lib/rfsa/adsp;/vendor/dsp;/system/lib/rfsa/adsp;/dsp"
export QWEN_HTP_PERFORMANCE=1
export QWEN_HTP_SHARED_IO=1
chmod +x qwen_runner
./qwen_runner "$PWD/runtime" "$PWD" "$PWD/input.f32" auto 512
The last stdout line is JSON with text, raw, language, token IDs, timing fields and token_limit. The encoder, decoder blocks and output head execute on HTP. CPU work includes Mel computation, tensor packing, embeddings, KV updates and token selection. Media decoding and VAD belong to the calling application.
English and Chinese short-sample generated token sequences matched the upstream BF16 Hugging Face greedy reference, excluding the reference's terminal EOS. A separate Android service business test completed 8/8 requests, processing 754.137 seconds of audio in 699.951 seconds, or 1.077× realtime throughput including admission and drain. Its p95 request latency was 538.80 seconds with multiple queued requests. These measurements are from the complete application, not a runner-only benchmark.
This build is useful for experimenting with on-device offline file transcription. VALIDATION.md explains the sample coverage, memory accounting, quality metrics and thermal measurements. The tested explicit prompts include Japanese and Portuguese, but recognition quality was evaluated only on the English/Chinese samples.
The 1024-position cache is shared by audio/prompt and generated tokens. max_tokens=512 is an upper bound and can be reduced by remaining cache capacity. Check token_limit and split long recordings in the calling application. The standalone runner processes one PCM input and provides no VAD, media-file decoding, automatic long-file splitting, streaming endpoint or timestamp alignment.
The separately tested Android application used continuous VAD and approximately 30-second chunks with a maximum 35-second model input. That application behavior is described in the validation notes; it is not a feature of this standalone runner. Compatibility with other chipsets or QAIRT versions has not been established. Timestamp generation requires application segmentation or a separate alignment model.
Model derivatives and included conversion/runtime source are provided under Apache-2.0; see LICENSE and NOTICE. The base model is published by the Qwen team. Conversion relies on Hugging Face Transformers and Qualcomm QAIRT/QNN. Qualcomm headers, runtime libraries and DSP skeletons are acquired separately and governed by their applicable terms.
audio_window.bin 355 MBoutput_head.bin 312 MBdecoder_0_s16.bin 127 MBdecoder_1_s16.bin 127 MBdecoder_2_s16.bin 127 MBdecoder_3_s16.bin 127 MBdecoder_4_s16.bin 127 MBdecoder_5_s16.bin 127 MBdecoder_6_s16.bin 127 MBdecoder_0_s1.bin 127 MBdecoder_1_s1.bin 127 MBdecoder_2_s1.bin 127 MBdecoder_3_s1.bin 127 MBdecoder_4_s1.bin 127 MBdecoder_5_s1.bin 127 MBdecoder_6_s1.bin 127 MBaudio_cnn.bin 23 MBtoken_bytes.bin 1 MB有文件?提交种子。
提交种子Qwen/Qwen3-ASR-0.6B-hf