What changed in every version, how we measured it, what didn't work and where we got it wrong. Every number is reproducible from the open weights and benchmark.
The newest Uzbek ASR models move to two release repositories, UzASR-CTC-600M and UzASR-CTC-220M, in safetensors, and the inference script can now say who spoke when.
We replayed the uzasr-large training recipe on the 220M model. Error fell from 10.5% to 9.4% WER on benchmark v3, decoding got faster, and the small model gained twice as much as the large one.
Error analysis showed many 'model errors' were mistakes in the reference transcripts. A native speaker listened to 374 disputed spots; 144 clips were corrected. All models were re-scored, with the bias disclosed.
How we cut Uzbek ASR error from 11.5% to 10.8% WER and made decoding 40% faster: spontaneous-speech training, an agreement-filtered YouTube corpus, a larger KenLM, and the experiments that didn't help.
We found that 20% of our Uzbek ASR benchmark had leaked into the KenLM text corpus. Here is how we found it, measured it and fixed it, and why honest WER is 11.5%, not 10.2%.
Adding a 4-gram KenLM language model to CTC decoding cut WER without retraining. This release was withdrawn a day later — the LM corpus overlapped the benchmark.
Our first public Uzbek ASR models: why Whisper fell short on spontaneous Uzbek, how we fine-tuned GigaAM-Multilingual, and how we measured it — 12.2% WER on a 2-hour benchmark.
