Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/QwenLM/Qwen3-ASR/llms.txt

Use this file to discover all available pages before exploring further.

Fine-tuning Qwen3-ASR lets you adapt the model’s speech recognition to your specific domain, acoustic environment, or language variety. The finetuning/qwen3_asr_sft.py script wraps Hugging Face Trainer and handles data loading, audio preprocessing, label masking, and checkpoint management — so you only need to supply a JSONL file and a handful of flags.

​
When to Fine-Tune

Out of the box, Qwen3-ASR covers 52 languages and dialects and performs competitively against the strongest commercial APIs. Fine-tuning is worthwhile when:
  • Domain-specific vocabulary — medical, legal, financial, or technical terms not well-represented in pre-training data.
  • Accent or dialect adaptation — recordings from a speaker population whose accent differs substantially from the training distribution.
  • Low-resource languages or dialects — languages where the pre-trained model has limited coverage.
  • Proprietary or stylized speech — internal meeting recordings, call-centre audio, or other controlled audio sources where you hold transcripts.
  • Consistent output formatting — you want the model to always follow a fixed punctuation or capitalisation style.
Fine-tuning modifies the existing weights rather than training from scratch, so even a few hundred labelled utterances can yield a meaningful improvement over the base model.

​
Prerequisites

Install the required packages before running the fine-tuning script.
1

Install core packages

pip install -U qwen-asr datasets
2

Install FlashAttention 2 (recommended)

FlashAttention 2 reduces GPU memory usage and speeds up training. It requires compatible hardware and a model loaded in torch.float16 or torch.bfloat16.
pip install -U flash-attn --no-build-isolation
If your machine has fewer than 96 GB of RAM but many CPU cores, limit the number of parallel compile jobs:
MAX_JOBS=4 pip install -U flash-attn --no-build-isolation
See the FlashAttention repository for full hardware compatibility details.
3

Prepare your JSONL training data

You need at least one JSONL file of audio-transcript pairs. An optional evaluation JSONL file can be supplied to monitor validation loss during training. See the Data Format page for the exact schema.

​
Fine-Tuning Workflow

The end-to-end process covers two main steps: formatting your data and launching the training script.

Data Format

Learn the JSONL schema, the language {Language}<asr_text>{transcript} text field format, and best practices for preparing WAV files.

Training

Run the fine-tuning script on a single GPU or scale to multiple GPUs with torchrun. Covers all CLI flags, checkpoint resumption, and a one-click shell script.

​
Quick Inference After Fine-Tuning

Once training completes, each saved checkpoint is a fully self-contained model directory that can be loaded directly with Qwen3ASRModel.from_pretrained. The script copies all required tokenizer and config files into the checkpoint folder automatically. 
import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "qwen3-asr-finetuning-out/checkpoint-200",
    dtype=torch.bfloat16,
    device_map="cuda:0",
)
results = model.transcribe(audio="path/to/audio.wav")
print(results[0].language)
print(results[0].text)
Replace "qwen3-asr-finetuning-out/checkpoint-200" with the path to whichever checkpoint you want to evaluate. The checkpoint number corresponds to the global training step at which it was saved.

Build docs developers (and LLMs) love

Get started for freeTalk to us