Uzbek speech recognition with Whisper-medium
Fine-tuned Whisper-medium on the 52-hour FeruzaSpeech corpus — 22,432 samples after augmentation — to 9.64% word error rate and 1.60% character error rate on the test split.
The part worth reusing is the pipeline rather than the checkpoint: text normalisation, Cyrillic-to-Latin handling, deduplication, duration filtering, and speed, pitch and MUSAN-noise augmentation, reproducible end to end.
Checkpoint: rustambekurokov/whisper-small-uzbek. Code: whisper-uzbek-fine-tuned.