Base Model: OpenAI Whisper Medium (769M parameters)
Language: Uzbek (uz)
Training Data: ~1,600 hours of Uzbek audio
Precision: BF16
Script: Latin (handles Russian loanwords in Latin script: "brat", "davay", "prosto", etc.)

Evaluation Results

Category	WER
Overall	16.7%
Clean Speech	~6-11%
Noisy/Augmented	~12-24%
Dialects	~16-25%

Evaluated on 1,864 samples across 8 diverse test sets.

Usage

Using Transformers

python

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import librosa

processor = WhisperProcessor.from_pretrained("Kotib/uzbek_stt_v1")
model = WhisperForConditionalGeneration.from_pretrained("Kotib/uzbek_stt_v1")

audio, sr = librosa.load("audio.wav", sr=16000)
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features

predicted_ids = model.generate(input_features, language="uz", task="transcribe")
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

Using Pipeline

python

from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="Kotib/uzbek_stt_v1",
    chunk_length_s=30,
    device="cuda"
)

result = pipe("audio.wav", generate_kwargs={"language": "uz", "task": "transcribe"})
print(result["text"])

Training

Trained in 3 stages using curriculum learning:

Stage	Hours
Foundation	725h
Robustness	394h
Domain Adaptation	474h

Intended Use

Uzbek speech-to-text transcription
Voice assistants and dictation
Media transcription and subtitling

Limitations

Performance degrades on very noisy audio
May struggle with heavy code-switching
Optimized for Uzbek only

License

Apache 2.0