NVIDIA Uzbek FastConformer, fine-tuned on the Uzbek Speech Corpus

Fine-tuned from nvidia/stt_uz_fastconformer_hybrid_large_pc (revision cd32f63d620a7b0cb461ed00ecfc45c852eb0bef) on a 25-hour, speaker-based subset of the Uzbek Speech Corpus (USC, ISSAI) train split, after our cleaning steps. Both heads of the hybrid model (transducer and CTC) were trained together; the start model's tokenizer is unchanged. Part of a university course project (COE025, Project 3). Code: https://github.com/ata-sesli/nlp-project-3

Training: 27144 clips, 23.922 h; learning rate 0.0004 (chosen on dev); best epoch 21 of 27 by WER on a fixed sample of 1000 dev clips, where WER went from 43.21% (start model) to 11.04%. Output text is lowercase with no punctuation (like our training text). License: see the start model's card.

Results (greedy decoding, no language model)

Model Set WER CER
transducer head (C4a) USC dev 11.60% 3.93%
CTC head (C4b) USC dev 12.46% 3.78%

WER/CER are on normalized text. Official numbers: benchmark/results/scores.csv in the code repo.

Usage

import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download

model = nemo_asr.models.ASRModel.restore_from(hf_hub_download("Waxech/stt-uz-fastconformer-usc", "stt_uz_fastconformer_usc_ft.nemo"))
print(model.transcribe(["clip.wav"])[0])          # transducer head (default)
model.change_decoding_strategy(decoder_type="ctc")
print(model.transcribe(["clip.wav"])[0])          # CTC head
Downloads last month
5
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Waxech/stt-uz-fastconformer-usc

Finetuned
(3)
this model

Dataset used to train Waxech/stt-uz-fastconformer-usc