Qwen3-ASR-1.7B GGUF

Qwen3-ASR-1.7B GGUF

Turns speech, singing and songs into text, or into English, with quantized GGUF files of Qwen3-ASR-1.7B for CrispASR. For people who want local transcription across 30 languages and 22 Chinese dialects.

Made by cstr on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Quantized GGUF files of Qwen3-ASR-1.7B, a speech recognition model, prepared for the CrispASR command-line tool. The model handles 30 languages and 22 Chinese dialects, and it copes with speech, singing and songs with background music. Audio inputs can run up to 5 minutes. The repository only changes the file format and number precision, with no new training.

Who it suits: People who want to transcribe or translate audio on their own computer and are comfortable running a command line.

Industry | Who uses it | How they use it | Example
Media | Podcasters and video editors | Turn recordings into text for captions and notes | A transcript of a short interview clip
Software | Developers adding speech input | Call a local command from a script | Transcribe a folder of audio files in a loop
Localization | Translators | Transcribe and translate a recording into English in one step | An English text from a Japanese clip
Music | Lyric collectors | Transcribe singing voice and songs with background music | Draft the lyrics of a track

You can skip it if your recordings run longer than 5 minutes without splitting, or if you prefer a graphical app over a command line.

Where to use it: The files are made for CrispASR, which you run from a terminal.

• CrispASR, the command-line tool these files are built for
• Plain CPU machines, since the card's speed test used four CPU threads
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Transcription | Podcasters | Text from speech, singing or songs with background music
Translation | Translators | English text from a recording in another language
Local batch jobs | Developers | Transcripts from a command line with no cloud service

How to set it up: Install CrispASR first, then either let it fetch the model on first use or download the files yourself. The Q4_K file is the card's pick for most people.

• CrispASR, the tool that runs the files
• Python, only if you use the Hugging Face download commands
• About 1.5 GB of disk space for the Q4_K file, up to 4.7 GB for F16 (about 5 GB of RAM)

1. Install CrispASR from the project the card names.
2. Optionally download the files with the Hugging Face tools.
3. Choose a quant: Q4_K for most uses, Q8_0 for more quality, F16 for full precision.
4. Run the crispasr command on your audio file.

pip install -U huggingface_hub
hf download cstr/qwen3-asr-1.7b-GGUF

How to use it: Point crispasr at an audio file with a model name or a model file. You can add a language hint, or switch on translation mode to get English text back.

# Auto-download and transcribe (Q4_K)
crispasr -m qwen3-1.7b -f audio.wav

# Or specify the model path directly
crispasr -m qwen3-asr-1.7b-q4_k.gguf -f audio.wav --backend qwen3

# With language hint
crispasr -m qwen3-1.7b -f audio.wav -l ja

# Translation mode
crispasr -m qwen3-1.7b -f audio.wav --translate --target-lang en

Tool | What it does
-m | Chooses the model, by short name or by file path
-f | Names the audio file to transcribe
-l | Gives a language hint, such as ja for Japanese
--translate with --target-lang | Turns on translation mode, with en as the target in the card's example
--backend qwen3 | Names the model family when you give a file path

Example: Run the translation command on a Japanese recording named audio.wav with the target language set to en. CrispASR reads the audio and prints the English text instead of the Japanese words.

For newcomers and experts: A speech recognition model turns spoken audio into written text, and a GGUF file is a compact copy that runs on an ordinary computer. A first small try is to run the first crispasr command on a short recording, since it fetches the model by itself.

For experts, the stack is a 24-layer Whisper-style audio encoder, a 2-layer MLP projector and a 28-layer Qwen3 decoder. The Q4_K file squeezes the language body but keeps the audio encoder at Q8_0, because a lower-bit encoder drifted and produced empty or repeated transcripts. On the card's JFK sample the encoder output matched the F16 reference at a minimum cosine of 0.9989 after the fix, up from 0.9632. The price is a file about 0.15 GB larger.

Audio file goes in
Encoder turns sound into features
Projector maps them for the decoder
Qwen3 decoder writes the text
Transcript or English translation comes out

Good to know: The three files differ in size and in how much memory they want.

Title: Approximate file size by quant
Q4_K: 1.5 GB
Q8_0: 2.5 GB
F16: 4.7 GB

• Origin: this repository only converts and quantizes the upstream model, so the upstream authors stay its providers.
• Fix: the Q4_K file was rebuilt after an earlier export sometimes returned blank text or looped a phrase. The other two files did not have that fault.
• Speed: on an 11-second English sample the Q4_K file ran at 0.2x realtime on CPU with 4 threads, and the card reports a perfect transcript.
• License: Apache 2.0, the same as the upstream model.
• Cost to run: none named, no API key, account or paid service.
• Data it reaches: it reads the audio files you point it at, and the first auto-download fetches the model.
• Not for: audio over 5 minutes in one go. Questions about training data go to the upstream model card.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Parameters: 2.3B
• Weights: about 1.2 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 11,523
• Likes: 17
• License: apache-2.0
• Languages: en, zh, ja, ko, de, fr, es, pt
• Based on: Qwen/Qwen3-ASR-1.7B
• Tags: gguf, speech-recognition, qwen3, automatic-speech-recognition

Links

• Model: https://huggingface.co/cstr/qwen3-asr-1.7b-GGUF
• Files: https://huggingface.co/cstr/qwen3-asr-1.7b-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download cstr/qwen3-asr-1.7b-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.