README.md

아직 제작중입니다.

Parrotron Speech-to-Speech

Strategy	Dataset	loss	WER
E0	Librispeech_test-clean	?
E1	Librispeech_test-clean	?

• E0 : Parrotron_no_ASR_decoder
• E1 : Parrotron

Parrotron Intro

Pytorch implementation of "Parrotron: An End-to-End Speech-to-Speech Conversion Model and its Applications to Hearing-Impaired Speech and Speech Separation".

Model

Feature

• Input wav : 80-dim log-Mel

  parameter	value
  N_FFT	1024
  Frame length	20ms
  Frame shift	12.5ms
  Window function	hann window
  fmin	125
  fmin	7600

• Targets wav : 1025-dim STFT

  parameter	value
  FFT	2048
  Frame length	20ms
  Frame shift	12.5ms
  Window function	hann window

Architecture

Print Model

Data

Dataset information

Train data set : train-clean-100 Test data set : test-clean

LibriSpeech ASR corpus : https://www.openslr.org/12

Data format

• 음성 데이터 : 16bit, mono 16k sampling WAV audio
• 정답 스크립트 :

    잠시보류

Dataset folder structure

• DATASET-ROOT-FOLDER

|--DATA
   |--train
      |--wav
         +--a.wav, b.wav, c.wav ...
      |--TTS_wav
         +--a.wav, b.wav, c.wav ...
      |--txt
         +--a.txt, b.txt, c.txt ...
   |--test
      |--wav
         +--a_test.wav, b_test.wav, c_test.wav ...
      |--TTS_wav
         +--a_test.wav, b_test.wav, c_test.wav ...
      |--txt
         +--a_test.txt, b_test.txt, c_test.txt ...

• train.csv, test.csv

  <wav-path>,<script-path>,<TTS_wav-path>
  460-172359-0076.wav,460-172359-0076.txt,460-172359-0076.wav
  460-172359-0077.wav,460-172359-0077.txt,460-172359-0077.wav
  460-172359-0078.wav,460-172359-0078.txt,460-172359-0078.wav
  460-172359-0079.wav,460-172359-0079.txt,460-172359-0079.wav
  460-172359-0080.wav,460-172359-0080.txt,460-172359-0080.wav
  ...
  

• english_unit.labels

  #id\char 
  0   _
  1   A
  2   B
  3   C
  ...
  26   Z
  27   '
  28    
  29   <s>
  30   </s>
  

References

Git hub References

• https://github.com/NVIDIA/tacotron2
• https://github.com/sooftware

Paper References

• Parrotron: An End-to-End Speech-to-Speech Conversion Model and its Applications to Hearing-Impaired Speech and Speech Separation (https://arxiv.org/abs/1904.04169)

computer power

• NVIDIA TITAN Xp * 4

Contacts

학부생의 귀여운 시도로 봐주시고 해당 작업에 대한 피드백, 문의사항 모두 환영합니다.

[email protected]로 메일주시면 최대한 빨리 답장드리겠습니다.