frozenlab/qwen3-asr

frozenlab/ qwen3-asr

2,380 Downloads Updated 1 month ago

vision 0.6b 1.7b

ollama run frozenlab/qwen3-asr

curl http://localhost:11434/api/chat \
  -d '{
    "model": "frozenlab/qwen3-asr",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='frozenlab/qwen3-asr',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'frozenlab/qwen3-asr',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

3 models

Size / Usage

Context

Input

qwen3-asr:latest

2.5GB · 64K context window · Text · 1 month ago

qwen3-asr:latest

2.5GB

64K

Text

qwen3-asr:0.6b

1.0GB · 64K context window · Text · 1 month ago

qwen3-asr:0.6b

1.0GB

64K

Text

qwen3-asr:1.7b

2.5GB · 64K context window · Text · 1 month ago

qwen3-asr:1.7b latest

2.5GB

64K

Text

Readme

Speech recognition models from the Qwen team (Alibaba), packaged for Ollama with the audio encoder (mmproj). 52 languages and dialects, automatic language detection, punctuation and capitalization. Weights: ggml-org GGUF conversion of Qwen3-ASR, Q8_0. License: Apache 2.0.

Tag	Size	Use
`1.7b` / `latest`	2.5 GB	best accuracy
`0.6b`	1.0 GB	fast, live transcription

Usage

Send the audio (16 kHz mono WAV, up to ~60 s per request) base64-encoded in the images field. Use an empty prompt:

curl http://localhost:11434/api/chat -d '{
  "model": "frozenlab/qwen3-asr",
  "stream": false,
  "messages": [{"role": "user", "content": "",
                "images": ["'"$(base64 -w0 speech.wav)"'"]}]
}'

The reply starts with the detected language, e.g. language German<asr_text>Hallo, wie geht's? — strip everything up to and including <asr_text>.

Measured (German conversation, 9.5 min, vs. human subtitles)

Model	Window	WER	Time per 30 s audio
1.7b	30 s	15.4 %	1.8 s
0.6b	30 s	19.3 %	1.0 s

Accuracy stays flat up to 60 s windows and degrades at ~2 min. Requires an Ollama version with qwen3a audio projector support.

Speech recognition models from the Qwen team (Alibaba), packaged for Ollama with the
audio encoder (mmproj). 52 languages and dialects, automatic language detection,
punctuation and capitalization. Weights: [ggml-org GGUF conversion](https://huggingface.co/ggml-org/Qwen3-ASR-1.7B-GGUF)
of [Qwen3-ASR](https://github.com/QwenLM/Qwen3-ASR), Q8_0. License: Apache 2.0.

| Tag | Size | Use |
|---|---|---|
| `1.7b` / `latest` | 2.5 GB | best accuracy |
| `0.6b` | 1.0 GB | fast, live transcription |

## Usage

Send the audio (16 kHz mono WAV, up to ~60 s per request) base64-encoded in the
`images` field. Use an **empty prompt**:

```bash
curl http://localhost:11434/api/chat -d '{
  "model": "frozenlab/qwen3-asr",
  "stream": false,
  "messages": [{"role": "user", "content": "",
                "images": ["'"$(base64 -w0 speech.wav)"'"]}]
}'
```

The reply starts with the detected language, e.g.
`language German<asr_text>Hallo, wie geht's?` — strip everything up to and including `<asr_text>`.

## Measured (German conversation, 9.5 min, vs. human subtitles)

| Model | Window | WER | Time per 30 s audio |
|---|---|---|---|
| 1.7b | 30 s | 15.4 % | 1.8 s |
| 0.6b | 30 s | 19.3 % | 1.0 s |

Accuracy stays flat up to 60 s windows and degrades at ~2 min. Requires an Ollama
version with `qwen3a` audio projector support.

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)