Uzbek Whisper Small STT v2
An Uzbek automatic speech recognition model based on Whisper Small, fine-tuned on Mozilla Common Voice 26 and converted to CTranslate2 FP16 for efficient inference with faster-whisper.
Evaluation
| Metric | Value |
|---|---|
| Training epochs | 3 |
| Optimization steps | 9,195 |
| Test samples | 12,405 |
| Test WER | 22.1379% |
| Test loss | 0.2795 |
Results were measured on the Uzbek test split of Mozilla Common Voice 26. Lower WER is better.
Installation
pip install faster-whisper
Usage
NVIDIA GPU
from faster_whisper import WhisperModel
model = WhisperModel(
"maqsudxo1ja/uz-whisper-small-stt-v2",
device="cuda",
compute_type="float16",
)
segments, info = model.transcribe(
"audio.wav",
language="uz",
beam_size=5,
)
text = "".join(segment.text for segment in segments).strip()
print(text)
CPU
from faster_whisper import WhisperModel
model = WhisperModel(
"maqsudxo1ja/uz-whisper-small-stt-v2",
device="cpu",
compute_type="int8",
)
Model Format
The main branch contains CTranslate2 model files and should be loaded with faster-whisper. It is not a standard Transformers checkpoint.
The previous Transformers-format revision is preserved in the backup-transformers-before-finetune-20260810 branch.
Limitations
Performance may vary across microphones, background noise, speaking styles, dialects, and audio domains that differ from Mozilla Common Voice. Validate the model on representative production audio before deployment.