
Base model	`nvidia/stt_uz_fastconformer_hybrid_large_pc`
Architecture	EncDecHybridRNNTCTCBPEModel (FastConformer encoder + RNNT/CTC decoders)
Training data	uzinfocom-edu-ai/uzbek-asr-curated-701h (~701 hours, 6 sources)
Parameters	~120M
Format	NeMo `.nemo`
Precision	fp32

Performance

Metric	Value
WER (RNNT)	16.01%
CER	6.12%
Test set size	27,142 utterances

WER by Source

Source	WER
uzbekvoice (read speech)	6.85%
news_youtube	28.61%
it_youtube	32.10%
podcasts_dialect	44.51%

Usage

python

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.from_pretrained(
    "uzinfocom-edu-ai/asr-uz-fastconformer-large"
)
model.eval()
transcriptions = model.transcribe(["audio.wav"], batch_size=1)
print(transcriptions[0].text)

Training Configuration

GPUs: 3× NVIDIA Tesla V100
Batch size: 8 × 3 GPUs × 8 gradient accumulation = 192 effective
Learning rate: 5e-5 (AdamW, cosine annealing, 5% warmup)
Epochs: 40 (best checkpoint at epoch 31)
Encoder freeze: first 3 epochs
Strategy: DDP (NCCL)

Data Sources

The model was trained on a curated multi-source Uzbek speech corpus (~701 hours):

Source	Hours	Domain
Common Voice (uz)	157.9h	Crowdsourced read speech
UzbekVoice	125.0h	Book/read speech (capped)
News YouTube	137.6h	Broadcast news
IT YouTube	134.8h	Tech content
Podcasts (Tashkent dialect)	94.8h	Conversational
Uzbek Speech Corpus (ISSAI)	50.5h	Studio recordings

Developed by

uzinfocom-edu-ai