【AI模型測試】使用Python實現語音檔案的特徵提取
阿新 • • 發佈:2020-11-19
參考地址:https://blog.csdn.net/qq_30091945/article/details/80941820
概述
語音識別是當前人工智慧的比較熱門的方向,技術也比較成熟,各大公司也相繼推出了各自的語音助手機器人,如百度的小度機器人、阿里的天貓精靈等。語音識別演算法當前主要是由RNN、LSTM、DNN-HMM等機器學習和深度學習技術做支撐。但訓練這些模型的第一步就是將音訊檔案資料化,提取當中的語音特徵。
MP3檔案轉化為WAV檔案
錄製音訊檔案的軟體大多數都是以mp3格式輸出的,但mp3格式檔案對語音的壓縮比例較重,因此首先利用ffmpeg將轉化為wav原始檔案有利於語音特徵的提取。其轉化程式碼如下:
from pydub import AudioSegment
import pydub
def MP32WAV(mp3_path,wav_path):
"""
這是MP3檔案轉化成WAV檔案的函式
:param mp3_path: MP3檔案的地址
:param wav_path: WAV檔案的地址
"""
pydub.AudioSegment.converter = "D:\\ffmpeg\\bin\\ffmpeg.exe"
MP3_File = AudioSegment.from_mp3(file=mp3_path)
MP3_File.export(wav_path,format="wav")
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
讀取WAV語音檔案,對語音進行取樣
利用wave庫對語音檔案進行取樣。程式碼如下:
import wave
import json
def Read_WAV(wav_path):
"""
這是讀取wav檔案的函式,音訊資料是單通道的。返回json
:param wav_path: WAV檔案的地址
"""
wav_file = wave.open(wav_path,'r')
numchannel = wav_file.getnchannels() # 聲道數
samplewidth = wav_file.getsampwidth() # 量化位數
framerate = wav_file.getframerate() # 取樣頻率
numframes = wav_file.getnframes() # 取樣點數
print("channel", numchannel)
print("sample_width", samplewidth)
print("framerate", framerate)
print("numframes", numframes)
Wav_Data = wav_file.readframes(numframes)
Wav_Data = np.fromstring(Wav_Data,dtype=np.int16)
Wav_Data = Wav_Data*1.0/(max(abs(Wav_Data))) #對資料進行歸一化
# 生成音訊資料,ndarray不能進行json化,必須轉化為list,生成JSON
dict = {"channel":numchannel,
"samplewidth":samplewidth,
"framerate":framerate,
"numframes":numframes,
"WaveData":list(Wav_Data)}
return json.dumps(dict)
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
繪製聲波折線圖與頻譜圖
程式碼如下:
from matplotlib import pyplot as plt
def DrawSpectrum(wav_data,framerate):
"""
這是畫音訊的頻譜函式
:param wav_data: 音訊資料
:param framerate: 取樣頻率
"""
Time = np.linspace(0,len(wav_data)/framerate*1.0,num=len(wav_data))
plt.figure(1)
plt.plot(Time,wav_data)
plt.grid(True)
plt.show()
plt.figure(2)
Pxx, freqs, bins, im = plt.specgram(wav_data,NFFT=1024,Fs = 16000,noverlap=900)
plt.show()
print(Pxx)
print(freqs)
print(bins)
print(im)
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
首先利用百度AI開發平臺的語音合API生成的MP3檔案進行上述過程的結果。
聲波折線圖
頻譜圖
全部程式碼
#!/usr/bin/python3
# -*- coding: utf-8 -*-
# @Time : 2018/7/5 13:11
# @Author : DaiPuwei
# @FileName: VoiceExtract.py
# @Software: PyCharm
# @E-mail :[email protected]
# @Blog :https://blog.csdn.net/qq_30091945
import numpy as np
from pydub import AudioSegment
import pydub
import os
import wave
import json
from matplotlib import pyplot as plt
def MP32WAV(mp3_path,wav_path):
"""
這是MP3檔案轉化成WAV檔案的函式
:param mp3_path: MP3檔案的地址
:param wav_path: WAV檔案的地址
"""
pydub.AudioSegment.converter = "D:\\ffmpeg\\bin\\ffmpeg.exe" #說明ffmpeg的地址
MP3_File = AudioSegment.from_mp3(file=mp3_path)
MP3_File.export(wav_path,format="wav")
def Read_WAV(wav_path):
"""
這是讀取wav檔案的函式,音訊資料是單通道的。返回json
:param wav_path: WAV檔案的地址
"""
wav_file = wave.open(wav_path,'r')
numchannel = wav_file.getnchannels() # 聲道數
samplewidth = wav_file.getsampwidth() # 量化位數
framerate = wav_file.getframerate() # 取樣頻率
numframes = wav_file.getnframes() # 取樣點數
print("channel", numchannel)
print("sample_width", samplewidth)
print("framerate", framerate)
print("numframes", numframes)
Wav_Data = wav_file.readframes(numframes)
Wav_Data = np.fromstring(Wav_Data,dtype=np.int16)
Wav_Data = Wav_Data*1.0/(max(abs(Wav_Data))) #對資料進行歸一化
# 生成音訊資料,ndarray不能進行json化,必須轉化為list,生成JSON
dict = {"channel":numchannel,
"samplewidth":samplewidth,
"framerate":framerate,
"numframes":numframes,
"WaveData":list(Wav_Data)}
return json.dumps(dict)
def DrawSpectrum(wav_data,framerate):
"""
這是畫音訊的頻譜函式
:param wav_data: 音訊資料
:param framerate: 取樣頻率
"""
Time = np.linspace(0,len(wav_data)/framerate*1.0,num=len(wav_data))
plt.figure(1)
plt.plot(Time,wav_data)
plt.grid(True)
plt.show()
plt.figure(2)
Pxx, freqs, bins, im = plt.specgram(wav_data,NFFT=1024,Fs = 16000,noverlap=900)
plt.show()
print(Pxx)
print(freqs)
print(bins)
print(im)
def run_main():
"""
這是主函式
"""
# MP3檔案和WAV檔案的地址
path1 = './MP3_File'
path2 = "./WAV_File"
paths = os.listdir(path1)
mp3_paths = []
# 獲取mp3檔案的相對地址
for mp3_path in paths:
mp3_paths.append(path1+"/"+mp3_path)
print(mp3_paths)
# 得到MP3檔案對應的WAV檔案的相對地址
wav_paths = []
for mp3_path in mp3_paths:
wav_path = path2+"/"+mp3_path[1:].split('.')[0].split('/')[-1]+'.wav'
wav_paths.append(wav_path)
print(wav_paths)
# 將MP3檔案轉化成WAV檔案
for(mp3_path,wav_path) in zip(mp3_paths,wav_paths):
MP32WAV(mp3_path,wav_path)
for wav_path in wav_paths:
Read_WAV(wav_path)
# 開始對音訊檔案進行資料化
for wav_path in wav_paths:
wav_json = Read_WAV(wav_path)
print(wav_json)
wav = json.loads(wav_json)
wav_data = np.array(wav['WaveData'])
framerate = int(wav['framerate'])
DrawSpectrum(wav_data,framerate)
if __name__ == '__main__':
run_main()