language:
- th
pipeline_tag: automatic-speech-recognition
library_name: transformers
base_model: Qwen/Qwen3-ASR-0.6B-hf
tags: - qwen3-asr
- thai
- speech-recognition
extra_gated_prompt: "Access is reviewed manually. Please describe your intended use."
extra_gated_fields:
Name: text
Affiliation: text
Intended use: text
qwen3-asr-0.6b-megaspeech-unfiltered-unrefined
Qwen/Qwen3-ASR-0.6B-hf fine-tuned for Thai ASR
(full fine-tuning).
Training data: MegaSpeech v3 unfiltered audios + unrefined texts (full original): gigaspeech-train, scraped, yodas (unfiltered), plus ai-cooking and thai-central (manually created, never filtered).
Clips longer than 30 s were skipped. Transcripts were used as-is (only whitespace collapsed).
Uploaded from: final
Results
| Eval set | Metric | Value |
|---|---|---|
| gigaspeech-dev (random 500 utterances, greedy) | CER (spaces removed) | 13.42% |
Usage
import librosa, torch
from transformers import AutoProcessor, Qwen3ASRForConditionalGeneration
repo = "ThuraAung1601/qwen3-asr-0.6b-megaspeech-unfiltered-unrefined"
processor = AutoProcessor.from_pretrained(repo)
model = Qwen3ASRForConditionalGeneration.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")
audio, _ = librosa.load("audio.wav", sr=16000)
inputs = processor.apply_transcription_request(audio=audio, language="th", return_tensors="pt").to(model.device, model.dtype)
out = model.generate(**inputs, max_new_tokens=448, do_sample=False)
print(processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
For audio longer than ~30 s, split it into chunks first; the model was trained on clips of 30 s or less.