Downloads · 30 days
41
39% of all-time downloads
Feelx8989/termux_os-asset-qwen3asr-htp
termux_os-asset-qwen3asr-htp is a automatic speech recognition model from Feelx8989. Use it when you need speech turned into text. The card lists the license as apache-2.0.
The pieces of Qwen/Qwen3-ASR-0.6B that Termux-OS builds itself: a mel front-end and an audio encoder as ONNX graphs for Qualcomm HTP, plus one custom decoder quantization.
Downloads · 30 days
41
39% of all-time downloads
All-time downloads
104
Public
Repo size
1.1 GB
Likes
0
Public
Click a slice to open those files.
.gguf727 MB · 66%
From the Hugging Face model README
The pieces of Qwen/Qwen3-ASR-0.6B that Termux-OS builds itself: a mel front-end and an audio encoder as ONNX graphs for Qualcomm HTP, plus one custom decoder quantization.
Termux-OS turns an Android phone into a local, open subsystem that ordinary users, developers and AI agents can extend together. Its Framework installs models as versioned, checksum-verified Assets fetched file by file from an immutable revision.
PCM ──► mel (ONNX, HTP) ──► encoder (ONNX, HTP) ──► decoder (GGUF, ggml-hexagon) ──► text
| Part | File | Origin | Hosted here |
|---|---|---|---|
| mel front-end | qwen3asr_mel.onnx | ours — a 16-node graph that also emits the per-window maximum | yes |
| audio encoder | qwen3asr_enc_tanh_fp16.onnx | ours — fp16 export with the clamp folded in | yes |
| decoder, Q4 | qwen3asr-q4out.gguf | ours — custom quantization, no public equivalent | yes |
| decoder, Q8 | Qwen3-ASR-0.6B-Q8_0.gguf | ggml-org/Qwen3-ASR-0.6B-GGUF — byte-identical, sha256 bca25981… | no, referenced |
⭐ The Q8 decoder is not re-hosted here. Its sha256 matches the official ggml build exactly, so Termux-OS references that repository instead. The test is not "is this file important" but "can this exact sha256 be found somewhere else" — re-hosting someone else's bytes means taking on responsibility for their integrity, licence and updates.
Q4 (727 MB) and Q8 (804 MB) are alternatives, not a set. A device installs one; the Termux Speech package declares them as optional assets so nothing is fetched until a tier is actually selected. The mel and encoder graphs are shared by both and are always required.
htp names the pipeline, not a device bindingThe ONNX graphs run through ONNX Runtime with the QNN execution provider; the GGUF decoder runs on ggml-hexagon. Both artefacts are portable. The compiled QNN EPContext is bound to the SoC, the QNN version and the model hash, so each device compiles its own and keeps it in that device's cache — it is never published here.
⚠ Two measured facts worth carrying: the mel stage is the part of this pipeline most affected by
thermal throttling, which is why it was moved onto the NPU at all; and the multimodal-projector path
(mmproj) produces numerically wrong results on ggml-hexagon and is usable on CPU only, so it is not
part of this pipeline.
Qwen3-ASR-0.6B is a work of the Qwen team, released under Apache-2.0. This repository redistributes derived artefacts under the same licence. The weights are unmodified except where a quantization is explicitly named as such; no claim is made over the original model.