quartznet-pytorch

Automatic Speech Recognition (ASR) on pytorch. Re-implementation on pytorch of Nvidia's Quartznet.

Features:

• Youtokentome tokenization with BPE dropout
• Augmentations: custom and audiomentations
• 3 datasets support: CommonVoice, Librispeech and LJSpeech
• Weights & Biases logging
• CTC beam search interation
• GPU-based MelSpectrogram

Trained models:

dataset	wer using dummy decoder	wer with ctc beam search	wer finetuned dummy decoder	wer finetuned ctc beam search
LJspeech	36.66	34.45	28.41	27.19

W&B Logs:

• CommonVoiceRU
• LJSpeech