automatic-speech-recognition audio

Online variant of pruned_transducer_stateless5 for Ukrainian: https://github.com/proger/icefall/tree/uk

Decoding demo using Sherpa: https://twitter.com/darkproger/status/1570733844114046976

Trained on pseudolabels generated by darkproger/pruned-transducer-stateless5-ukrainian-1 on the noisy 1200 hours training set. Common Voice data was used only for validation.

Tensorboard run

./pruned_transducer_stateless5/train.py \
  --world-size 2 \
  --num-epochs 31 \
  --start-epoch 1 \
  --full-libri 1 \
  --exp-dir pruned_transducer_stateless5/exp-uk-filtered2 \
  --max-duration 600 \
  --use-fp16 1 \
  --num-encoder-layers 18 \
  --dim-feedforward 1024 \
  --nhead 4 \
  --encoder-dim 256 \
  --decoder-dim 512 \
  --joiner-dim 512 \
  --bpe-model uk/data/lang_bpe_250/bpe.model \
  --causal-convolution True \
  --dynamic-chunk-training True