---
title: libwaifu-qwen3-asr-0.6b
canonical_url: "https://www.modelscope.cn/models/ling0322/libwaifu-qwen3-asr-0.6b"
md_url: "https://www.modelscope.cn/models/ling0322/libwaifu-qwen3-asr-0.6b.md"
repository: ling0322/libwaifu-qwen3-asr-0.6b
last_updated: 2026-11-09
license: apache-2.0
base_model:
  - Qwen/Qwen3-ASR-0.6B
base_model_relation: finetune
parameters: 782.4M
tensor_type:
  - F16
library_name:
  - safetensors
language:
  - zh
  - en
  - yue
  - ar
  - de
  - fr
  - es
  - pt
  - id
  - it
  - ko
  - ru
  - th
  - vi
  - ja
  - tr
  - hi
  - ms
  - nl
  - sv
  - da
  - fi
  - pl
  - cs
  - fil
  - fa
  - el
  - ro
  - hu
  - mk
downloads: 2
stars: 0
tags:
  - libwaifu
  - qwen3-asr
  - speech-recognition
---

# libwaifu-qwen3-asr-0.6b

> libwaifu-qwen3-asr-0.6b - ling0322 在 ModelScope 开源的模型。Qwen3-ASR-0.6B — the libwaifu package

ling0322/libwaifu-qwen3-asr-0.6b 是 ModelScope 魔搭社区上的 782.4M 参数机器学习模型，采用 apache-2.0 许可，基于 Qwen/Qwen3-ASR-0.6B 构建。

- **Repository**: ling0322/libwaifu-qwen3-asr-0.6b
- **License**: apache-2.0
- **Parameters**: 782.4M
- **Base model**: Qwen/Qwen3-ASR-0.6B
- **Tags**: libwaifu, qwen3-asr, speech-recognition
- **Downloads**: 2
- **Stars**: 0
- **Last updated**: 2026-11-09

Source: https://www.modelscope.cn/models/ling0322/libwaifu-qwen3-asr-0.6b

---

# Qwen3-ASR-0.6B — the libwaifu package

**Qwen3-ASR-0.6B**, converted to the package format
[libwaifu](https://github.com/ling0322/libwaifu) reads. A recording in; what was said in it, and in
which language, out -- thirty languages, told which or left to find out, with a sentence of
context to help it hear names. The same network; a different file layout, one manifest, and the
weights in half precision.

This is a **modified copy** under the original Apache License 2.0 — see `LICENSE` and `NOTICE` for
what was changed. It is **not an official Qwen product and is not endorsed by them**. The original
weights are at [Qwen/Qwen3-ASR-0.6B](https://huggingface.co/Qwen/Qwen3-ASR-0.6B).

## What is here

| file | on disk | what it is |
|---|---|---|
| `qwen3-asr.yaml` | 1.2 KB | the manifest: what the model is and which files it is made of |
| `qwen3-asr.safetensors` | 1.56 GB | 782 M parameters in float16: the audio tower and Qwen3-0.6B's decoder |
| `qwen3-asr.tokenizer.json` | 11 MB | the Qwen vocabulary, with the audio markers and `<asr_text>` |

The release is bfloat16; every value in half's range survives the narrowing whole, and the
exporter refuses a tensor with anything past half's largest. The text head is left out: in the
release it is the embedding table bit for bit, and the embedding is used in its place.

## How close it is

Checked against a float32 reference put together from transformers' `WhisperFeatureExtractor`,
upstream's audio encoder written out in plain torch, and transformers' `Qwen3ForCausalLM` with the
release's weights, on three real recordings (RTX 5060 Ti):

| | zh, 3.5 s | 13.75 s at 22.05 kHz | en, 5.3 s |
| --- | --- | --- | --- |
| mel, relative error | 1.4e-5 | 1.5e-5 | 1.7e-5 |
| encoder, relative error | 0.18% | 0.49% | 0.24% |
| decoder, worst log probability | 0.037 | 0.039 | 0.059 |
| tokens | the same 14 | the same 39 | the same 19 |

The package reads in 0.6 s and writes the 13.75 s recording down in 0.36 s. The reference tensors
are `qwen3-asr_test.safetensors` in
[ling0322/libwaifu_test_data](https://huggingface.co/ling0322/libwaifu_test_data).

## Transcribing with it

```bash
cargo run --release -p waifu-tools --bin transcribe -- qwen3-asr.yaml speech.wav
cargo run --release -p waifu-tools --bin transcribe -- qwen3-asr.yaml speech.wav English
```

```rust
let asr = Qwen3Asr::from_manifest(Device::Cuda, u64::MAX, &manifest)?;
let said = asr.transcribe(&recording, &TranscriptionOptions::default(), &mut report)?;
// Some(Transcript { text: "希望你以后能够做得比我还好哟。", language: "Chinese" })
```

In the model manager it is `qwen3-asr`. `docs/qwen3_asr.md` in the libwaifu repository has the
whole pipeline, how it was checked, and how this package was exported.

## Where this package knowingly differs from upstream

- It runs in half on the GPU and on the processor; upstream runs bfloat16.
- The audio encoder keeps attention inside its eight-second windows, as upstream's FlashAttention
  and vLLM paths do; upstream's transformers backend under sdpa or eager attends across the whole
  recording.
- No timestamps (upstream's Qwen3-ForcedAligner is not ported), no streaming, no batching.

## Attribution

Qwen3-ASR-0.6B is licensed under the Apache License 2.0 — see
https://huggingface.co/Qwen/Qwen3-ASR-0.6B. See `NOTICE` for exactly what was changed.