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.

The fine-tuning script expects training data as a JSONL file — one JSON object per line — where each line maps a WAV audio file to its ground-truth transcript. Preparing this file correctly is the most important step before you start training, because the text field encoding directly controls whether the model learns language identification alongside transcription.

​
JSONL Format Specification

Each line in the training file must be a valid JSON object containing exactly two fields:
FieldTypeDescription
audiostringAbsolute or relative path to a WAV audio file on disk
textstringTranscript string, including a required language prefix (see below)
The file must use the .jsonl extension (or .json for a JSON-lines file). One record per line; no trailing commas; UTF-8 encoding.
The script uses the Hugging Face datasets library to load the JSONL file. Any additional fields present in each line are silently dropped during preprocessing.

​
The text Field Format

The text value is not a bare transcript. It follows a structured template that mirrors the model’s output format: 
language {Language}<asr_text>{Transcript}
The <asr_text> token is a fixed delimiter — the model was trained with this separator and expects it at inference time as well.

​
With Known Language

When you know the spoken language, supply it as an English-capitalised name immediately after language : 
language English<asr_text>This is a test sentence.
language Chinese<asr_text>你好世界
language French<asr_text>Bonjour le monde
language Japanese<asr_text>こんにちは世界
Training with an explicit language label teaches the model to produce correct language identification output for that utterance.

​
Without Language Information

If you do not have reliable language labels for your data, use the literal keyword None: 
language None<asr_text>Transcript text here
When language None is used, the model will not learn language-detection behaviour from that example — only the transcription path is trained. Use this when your language labels are unavailable or unreliable rather than guessing the language, which could introduce noisy supervision.

​
Example JSONL File

Below is a representative training file combining multiple languages and the None fallback: 
{"audio":"/data/wavs/utt0001.wav","text":"language English<asr_text>This is a test sentence."}
{"audio":"/data/wavs/utt0002.wav","text":"language Chinese<asr_text>你好世界"}
{"audio":"/data/wavs/utt0003.wav","text":"language None<asr_text>transcript without language info"}
{"audio":"/data/wavs/utt0004.wav","text":"language English<asr_text>Another example from the training set."}
{"audio":"/data/wavs/utt0005.wav","text":"language French<asr_text>Une autre phrase en français."}
Each line is independent. You can mix languages freely within the same file.

​
Language Prefix Options

The {Language} placeholder should be a capitalised English language name. Examples of valid values:
  • English, Chinese, French, German, Spanish, Japanese, Korean
  • Cantonese, Arabic, Hindi, Portuguese, Russian, Thai, Vietnamese
  • Any of the 52 languages and dialects supported by Qwen3-ASR
  • None — when language information is unavailable
Using language None does not harm transcription quality. It simply prevents that sample from contributing to the language-identification head’s gradient update.

​
Preparing Audio Files

The data collator inside qwen3_asr_sft.py loads each WAV file at runtime using librosa and resamples to 16,000 Hz mono. Requirements for your audio files:
  • Format: WAV (uncompressed PCM is safest; librosa can handle most sub-formats)
  • Sample rate: any — the loader resamples to 16 kHz automatically
  • Channels: any — the loader forces mono
  • Paths: must be accessible from the machine running the training script; absolute paths are the safest option
Missing or unreadable audio files cause a runtime error during the collation step. Verify that every path in your JSONL file resolves correctly before starting a long training run.

​
Optional Evaluation File

You can pass a second JSONL file with the --eval_file flag. It must follow the same format as the training file. When provided, the script evaluates on this split every --save_steps steps and logs the validation loss alongside the training loss.

Build docs developers (and LLMs) love

Get started for freeTalk to us