ss87021456 / mfcc_ctc_speech

apply mfcc feature of waveform with the LSTM + CTC loss architecture

mfcc_ctc_speech

Using CTC loss function combine mfcc feature and LSTM architect
test on youtube dataset

Dependency:

• for create label (gen_label.py)
  python3 - 3.6.1
  webvtt-py - 0.4.0
• for cuting dataset (mp4_to_cut_wav.py)
  python2 - 2.7.14
  moviepy - 0.2.3.2
  cv2 - 3.3.0
• for training (ctc_speech_recognition.py)
  python2 - 2.7.14
  tensorflow - 1.4.0

Usage:

Step 1 : Download youtube vedio with cc subtitle
Step 2 : python3 gen_label.py - to generate clear label
Step 3 : python mp4_to_cut_wav.py - to generate wav dataset
Step 4 : python ctc_speech_recognition.py - training

Dataset description:

Training Process: