feats.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

from email import header
import os
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.style as ms
import pandas as pd
import librosa
import librosa.display

class FeatExtrct():
  def __init__(self, data, sr):
    self.frame_length = 0.04 # 40ms
    self.frame_shift = 0.01 # 30ms
    self.win_length = int(self.frame_length*sr)
    self.hop_length = int(self.frame_shift*sr)
    self.n_fft = 256
    self.data = data
    self.sr = sr
  
  def _pitch_extract(self):
    pitches, magnitudes = librosa.core.piptrack(y=self.data, sr=self.sr, \
      win_length=self.win_length, hop_length=self.hop_length, \
      threshold=1, fmin=75, fmax=1600)
    max_indexes = np.argmax(magnitudes, axis=0)
    pitches = pitches[max_indexes, range(magnitudes.shape[1])]
    return pitches
  
  def _mfcc_extract(self):
    mfcc = librosa.feature.mfcc(y=self.data, sr=self.sr, \
           win_length=self.win_length, hop_length=self.hop_length, \
           center=True, n_mfcc=13)
    mfcc_delta = librosa.feature.delta(mfcc, mode="nearest")
    mfcc_delta2 = librosa.feature.delta(mfcc, mode="nearest", order=2)
    mfcc_delta12 = np.vstack((mfcc, mfcc_delta, mfcc_delta2))
    return mfcc_delta12
  
  def _zero_crossing_rate(self):
    zcr = librosa.feature.zero_crossing_rate(self.data, \
              frame_length=self.win_length, hop_length=self.hop_length)
    return zcr
  
  def get_feats(self):
    pitch = self._pitch_extract()
    mfcc = self._mfcc_extract()
    zcr = self._zero_crossing_rate()
    return np.vstack((pitch, mfcc, zcr))


class DataPrep():
  def __init__(self, textgrid):
    self.textgrid = textgrid
    self.wavlist = None
  
  def _get_all_wavs(self):
    df = pd.read_csv(self.textgrid, header=0, sep=",")
    self.wavlist = df.loc[:,["文件名", "序号", "音节起始时刻点", "韵母起始时刻点", "音节结束时刻点"]]
    # print(self.wavlist)
  
  def _get_label(self, labelfile):
    with open(labelfile, 'r') as fp:
      return []
  
  def _get_feats(self, data, sr):
    FE = FeatExtrct(data, sr)
    feats = FE.get_feats()
    return feats

  def write_feat(self, feat_pa):
    self._get_all_wavs()
    for i in range(len(self.wavlist)):
    # for i in range(0,1):
      conso_offset = self.wavlist.loc[i, '音节起始时刻点']
      conso_duration = self.wavlist.loc[i, '韵母起始时刻点'] - self.wavlist.loc[i, '音节起始时刻点']
      vowel_offset = self.wavlist.loc[i, '韵母起始时刻点']
      vowel_duration = self.wavlist.loc[i, '音节结束时刻点'] - self.wavlist.loc[i, '韵母起始时刻点']
      wav_path = self.wavlist.loc[i, "文件名"]
      person_name = wav_path.split("/")[-1].split(".")[0] + "_" + str(self.wavlist.loc[i, "序号"])
      print("person_name : ", person_name)

      conso_data, _ = librosa.load(wav_path, sr=16000, offset=conso_offset, duration=conso_duration)
      vowel_data, _ = librosa.load(wav_path, sr=16000, offset=vowel_offset, duration=vowel_duration)
      final_feats = []
      # 辅音标签为0， 元音标签为1
      if conso_duration >= 0.05:
        conso_feats = self._get_feats(conso_data, 16000)
        conso_feats = np.r_[conso_feats, np.expand_dims(np.zeros(conso_feats[0].size),axis=0)]
        final_feats = conso_feats
      if vowel_duration >= 0.05:
        vowel_feats = self._get_feats(vowel_data, 16000)
        vowel_feats = np.r_[vowel_feats, np.expand_dims(np.ones(vowel_feats[0].size),axis=0)]
        if len(final_feats) == 0:
          final_feats = vowel_feats
        else:
          final_feats = np.hstack((final_feats, vowel_feats))
      # person_feats = np.hstack((conso_feats, vowel_feats))
      feat_path = feat_pa + "_" + person_name + ".npy"
      np.save(file=feat_path, arr=final_feats)

if __name__ == '__main__':
  wavdir = "data/textgrid/select_TextGrid.csv"
  feat_path = "data/textgrid/feat"
  DP = DataPrep(wavdir)
  DP.write_feat(feat_path)