Mega-ASR

Mega-ASR

Transcribes English and Chinese speech from noisy, echoey, clipped or overlapping recordings, and routes clean audio to the base model. For teams with difficult audio.

Made by zhifeixie 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: Mega-ASR is a speech recognition system built for badly degraded audio. It targets noise, echo, clipping, narrow bandwidth and overlapping voices, where ordinary systems often return nothing, skip words or invent text. It is built on Qwen3-ASR-1.7B and adds adaptation weights plus an audio quality router.

Who it suits: It suits teams that transcribe recordings made in poor conditions, such as far-field microphones, street noise or compressed audio.

Industry | Who uses it | How they use it | Example
Media | Transcription teams | Transcribe interviews recorded outdoors or on weak microphones | Getting text from a noisy street interview
Customer service | Call analysis teams | Transcribe compressed or noisy call audio | Reading calls with line noise
Meetings | Note takers | Transcribe far-field and echoey room recordings | A conference room recording with echo
Research | Speech researchers | Test recognition under compound distortions | Checking word error rates on degraded audio

You can skip it if your audio is always clean, since the base model already handles that case and the router sends such audio to it anyway.

Where to use it:

• Python on Windows, macOS or Linux, using the project's inference script or its Python class
• A conda environment with Python 3.10, which the card's setup uses
• Evaluation runs, since the project includes a word error rate script

Use case | Who | What they get back
Transcribe degraded audio | Transcription teams | Text from noisy, echoey or clipped recordings
Auto-pick the best path | Developers | A router that chooses the degraded-audio or base path per file
Measure accuracy | Researchers | Word or character error rates from a test file

How to set it up:

• Python 3.10 and conda
• The Mega-ASR code and its requirements file
• The downloaded checkpoint folder placed where the code expects it

1. Download the checkpoint with Hugging Face's own steps.
2. Get the Mega-ASR code and create a conda environment.
3. Install the requirements.
4. Put the checkpoint folder in the ckpt directory as the card says.

pip install -U huggingface_hub
hf download zhifeixie/Mega-ASR

conda create -n mega-asr python=3.10 -y
conda activate mega-asr
pip install -r requirements.txt

How to use it: You run the inference script with an audio file and the checkpoint folder. A second option turns routing off so every file takes the degraded-audio path.

python infer.py --audio /path/to/audio.wav --ckpt_dir ckpt/Mega-ASR

python infer.py --audio /path/to/audio.wav --ckpt_dir ckpt/Mega-ASR --routing false

In Python, the card gives this class-based call.

from MegaASR.model.megaASR import MegaASR
model = MegaASR(
    model_path="ckpt/Mega-ASR/Qwen3-ASR-1.7B",
    router_checkpoint="ckpt/Mega-ASR/audio_quality_router/best_acc_model.pt",
    routing_enabled=True,
)
result = model.infer("/path/to/audio.wav", return_route=True)
print(result)

Example: A person has a recording of an interview on a busy street. They run the inference script on the file. What comes back is the transcript text. With the route option in Python, they also see which path the router picked for that file.

For newcomers and experts: If you are new to this, run the script on one short clip and read the transcript. Keep routing on, which is the default, so clean clips go to the base model.

If you know the field, the release bundles the Qwen3-ASR-1.7B base, tuned weights and a small classifier that scores how damaged the audio is. Its cutoff is 0.5 unless you change it. Decoding is greedy and stops at 256 new tokens.

Load an audio file
The router scores the audio quality
Choose the degraded-audio or base path
Decode greedily up to 256 tokens
Return the transcript

Good to know:

• Limits: the card shows results as a chart and gives no number table in the part shown.
• Defaults: decoding is greedy, so temperature and sampling settings have no effect.
• Languages: the record lists English and Chinese.
• License: Apache 2.0.
• Cost to run: none named.
• Data it reaches: it copies model files to your computer and reads the audio files you give it.

On Hugging Face, 14 November 2026

• Task: automatic speech recognition
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 104
• License: apache-2.0
• Languages: en, zh
• Based on: Qwen/Qwen3-ASR-1.7B
• Quantized versions: 7
• Fine-tunes: 1
• Tags: safetensors, automatic-speech-recognition, speech-recognition, audio, robust-asr, qwen3-asr

Links

• Model: https://huggingface.co/zhifeixie/Mega-ASR
• Files: https://huggingface.co/zhifeixie/Mega-ASR/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download zhifeixie/Mega-ASR

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