GigaAM-Uzbek β v2 (gigaam-uzbek-v2.ckpt) and v1
Uzbek (Latin script) speech recognition: full fine-tunes of
GigaAM multilingual_ctc (220M Conformer-CTC).
Academic research release β every training source is listed above and cited below;
the model inherits the most restrictive input license (non-commercial).
| File | Model | Training data |
|---|---|---|
gigaam-uzbek-v2.ckpt |
E8 (recommended) | 776h gold + 357h filtered pseudo-labeled conversational speech (1,242h effective) |
gigaam-uzbek-v1.ckpt |
E3 | 646h gold (UzbekVoice + Common Voice) |
Results (WER %, greedy CTC, identical normalization)
| Benchmark | domain | zero-shot 220M | zero-shot 600M | v1 (E3) | v2 (E8) |
|---|---|---|---|---|---|
| UzbekVoice dev | read | 5.08 | 4.15 | 2.18 | 2.28 |
| Common Voice uz test | read | 7.67 | 5.77 | 7.52 | 6.88 |
| FLEURS uz test* | read | 10.47 | 7.52 | 11.55 | 8.71 |
| USC test | studio read | 9.72 | 7.17 | 13.17 | 8.98 |
| FeruzaSpeech testβ | audiobook/news | 6.08 | 5.42 | 6.44 | 2.32 |
| Spontaneous benchmark (rustam1221)β‘ | podcasts/interviews | 16.65 | 15.87 | 23.40 | 13.66 |
| Telegram voice messages (OvozifyLabs) | conversational | 22.53 | 20.45 | 30.75 | 23.93 |
| Qashqadaryo dialect podcast (Fdev071) | dialect | 49.93 | 48.62 | 57.78 | 51.81 |
* 17% of FLEURS references (digit rows) excluded β no number expansion. β FeruzaSpeech train split (same single speaker) is in v2 training data; score reflects speaker familiarity by corpus design. β‘ Caveat: 63 of 310 benchmark references share 8-word n-grams with the pseudo-labeled pools (the benchmark's source podcasts also appear in the islomov YouTube pools); a leak-guarded retrain is in progress. All other test sets show 0% n-gram overlap with training data.
Training (v2 / E8)
Fresh fine-tune from multilingual_ctc: lr 5e-5, 6 epochs, batch 64, bf16-mixed, SpecAugment
defaults, Common Voice Γ3 upsampled, checkpoint by WER on a 4.4k-utterance mixed dev set.
Pseudo-labels: Gemini transcripts kept only where multilingual_large_ctc greedy output agrees
within CER β€ 10% (yield: news 68%, IT 69%, Abduqayum 67%, Tashkent podcasts 32%).
Hardware: 1Γ NVIDIA B200, ~6.5h.
Normalization
Single normalizer for all training and reference text
(source):
NFC β apostrophe unification (Κ» ΚΌ β β Β΄ β¦ β ') **before** punctuation stripping β lowercase β GigaAMnormalize_raw_text` β charset {space, ', aβz}; Cyrillic sources transliterated (1995 Latin).
Usage
import gigaam
model = gigaam.load_model("gigaam-uzbek-v2.ckpt")
print(model.transcribe("audio.wav")) # 16 kHz mono
Limitations
Latin output only; no digits (number expansion pending); dialect speech still ~50% WER; non-commercial (inherits CC-BY-NC from one pseudo-label source).
Code, experiment registry, full report: https://github.com/SamariddinS/uzbek-voice
- Downloads last month
- -