Base model: Qwen/Qwen3-ASR-1.7B
Language: Uzbek (uz)
Training data: ~160K samples from FLEURS + USC + YouTube (IT, News, Podcasts)
Training: 5 epochs, lr=1e-5, continues from v1 checkpoint-2433

Usage

python

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "Gearnode/qwen3-asr-uzbek-v2",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    max_new_tokens=448,
    forced_aligner="Qwen/Qwen3-ForcedAligner-0.6B",
    forced_aligner_kwargs=dict(dtype=torch.bfloat16, device_map="cuda:0"),
)

results = model.transcribe(
    audio=[(audio_array, 16000)],
    language=["Uzbek"],
    return_time_stamps=True,
)
print(results[0].text)

Training Datasets

Google FLEURS uz_uz (2,943 samples)
Uzbek Speech Corpus (100,767 samples)
YouTube IT Uzbek (21,016 samples)
YouTube News Uzbek (20,795 samples)
YouTube Podcasts Uzbek (14,547 samples)

Evaluation

Model	Uzbek Quality	Notes
This model	Best	Some repetition handled by post-processing
Meta MMS (mms-1b-all)	Passable	Use lang code `uzb-script_latin`
Whisper large-v3	Poor	Hallucination loops on Uzbek
Base Qwen3-ASR	Poor	Mixed languages

License

Apache 2.0 (same as base model)