python nn 聲音辨識 -1 傅立葉轉換

辨識環境聲音判斷出這是甚麼聲音

主要想法是用頻域赫茲分布做判斷
首先先錄三種聲音的聲音檔，我每一種聲音重複錄製九個sample，所以有27個wav檔

讀取wav檔案，作正規劃處理，在做傅立葉轉換
一開始由於自己做的fft_domyself() function跟其他聲音處理軟體做出來的不一樣，數值非常小，之後找到有內建的spectrogram() function，使用也是數值很小，最後才想到要20log()，但由於內建的覺得不適合自由調整，所以還是修改自己的function。

import matplotlib.pyplot as plt
from scipy.fftpack import fft
from scipy.io import wavfile # get the api
import numpy as np
from scipy import signal
import math
import pyaudio
import wave
import time
#buf=[]
 
#讀入單一檔案測試
def init():
    fs, data = wavfile.read('0106.wav') # load the data
    b=[(ele/2**16.) for ele in data] # this is 16-bit track, b is now normalized on [-1,1)
    c = fft(b) # calculate fourier transform (complex numbers list)
    d = len(c)/2  # you only need half of the fft list (real signal symmetry)
    plt.subplot(211)
    plt.plot(b,'b')
    plt.title('time line')
    plt.subplot(212)
    plt.plot(abs(c[:(d-1)]),'r')
    plt.title('fft') 
    plt.show()
    return fs, b    #取樣率 data
 
#轉成頻域
def fft_domyself(b, fs, plotshow=None):
    global ftmap
    #set value
    one_second_block_num = 21*2                   #one second 擷取次數
    second = len(b)/44100
    second_len = int(one_second_block_num*second)
    fft_num = int(fs/one_second_block_num)        #one second 間格點數 #2100
    ftmap = np.zeros((int(fft_num/2),second_len)) #建立fft陣列的圖片大小  
    startime = 0#362000
 
    #轉成fft 並儲存在fft陣列
    for i in range(0,second_len):
        endtime = startime + fft_num
        ffttemp = abs(fft(b[startime:endtime]))
        ftmap[0:fft_num,i] = ffttemp[0:int(len(ffttemp)/2)]
        startime = endtime
 
    #做20log()
    for i in range(0, len(ftmap[:,1])):
        for j in range(0, len(ftmap[1,:])):
            try:
                #ftmap[i,j] = 20*math.log10(ftmap[i,j])
                ftmap[i,j] = 20*math.log(ftmap[i,j], 10)
            except:
                #由於 log(0) 輸入0會error
                print('log error with ftmap: ', ftmap[i,j])
        
    ftmap = ftmap + abs(np.min(ftmap))  
    lineft = np.median(ftmap, axis=1)   #用二維的方式，查看頻域的整體資料情況
    
    if plotshow:
        if plotshow == 'linefft':
            plt.plot(lineft, 'b')
            plt.show()
        else:
            
            plt.plot(lineft,'g') 
            plt.show()
            plt.imshow(ftmap)
            plt.xlabel('Frequency [Hz]')
            plt.ylabel('Time [sec]')
            plt.colorbar()
            plt.show()
     
    return ftmap, lineft
 
#使用內建的function將聲音轉成頻域
def spectrogram_(b, fs, plotshow=None):
    #x1 = np.array(b, dtype = float)
    f, t, Sxx = signal.spectrogram(b, fs)    #output 時頻譜
    
    for i in range(0, len(Sxx[:,1])):
        for j in range(0, len(Sxx[1,:])):
            Sxx[i,j] = 20*math.log10(Sxx[i,j])
            #print('do')
    Sxx = Sxx + abs(np.min(Sxx))
    
    if plotshow: 
        plt.pcolormesh(t, f, Sxx)             #draw 時頻譜
        plt.ylabel('Frequency [Hz]')
        plt.xlabel('Time [sec]')
        plt.colorbar()
        plt.show()
    return Sxx
if __name__ == '__main__':
    
    fs, b = init()
    fftmap = fft_domyself(b, fs, plotshow = True)   #自己做的
    sx = spectrogram_(b, fs, plotshow = True)       #內建function

2018/2/24更新(添加註解)

img

lineft

ftmap

spectrogram_()

讀多個聲音檔，並轉換成fft矩陣

我是用三個鈴鐺做的，無聲音檔就是沒有鈴鐺聲只有背景聲

  2018/2/24更新
輸出：
linef = [0~20000Hz頻域的大小 ,  0~20000Hz頻域的大小 ,  0~20000Hz頻域的大小 ......]
answer = [ [0.0 , 1.0 , 0.0 , 0.0] , [0.0 , 0.0 , 1.0 , 0.0] , [0.0 , 0.0 , 0.0 , 1.0] ......]
                [0.0 , 1.0 , 0.0 , 0.0] 代表是第一類聲音

def read_file(filename, plotshow=None):
    fs, data = wavfile.read(filename) # load the data
    b=[(ele/2**16.) for ele in data] # this is 16-bit track
    if plotshow:
        plt.plot(b,'b')
        plt.title('time line')
        plt.show()
    return b, fs
 
def read_to_mat():
    file_road = './train_data/'
    filehead = ['00', '01', '02', '03']
    linef = []
    answer = []
    for i in range(1, 10):                  #last two number
        for j in range(0, len(filehead)):   #first two number
            readURL = file_road + filehead[j] + '%02d'%(i) + '.wav'
            x = [0.0, 0.0, 0.0, 0.0]
            x[int(filehead[j])] = 1.0
            
            audio_value, fs = read_file(readURL)
            fftmap, linefft = fft_domyself(audio_value, fs)
            normalized_linefft = (linefft - np.mean(linefft)) / np.std(linefft)
            
            answer.append(x)
            linef.append(normalized_linefft.tolist())
            print(readURL, x)
    return linef, answer

錄製聲音

def recoder(recoder_second, filename = None, plotshow=None):
    NUM_SAMPLES  =  100       # pyAudio內部緩存的塊的大小
    SAMPLING_RATE = 44100     #取樣頻率
    #recoder_second = 1       
    '''
    LEVEL  =  1500             #聲音保存的閾值
    COUNT_NUM  =  20           # NUM_SAMPLES個取樣之內出現COUNT_NUM個大於LEVEL的取樣則記錄聲音
    SAVE_LENGTH  =  8          #聲音記錄的最小長度：SAVE_LENGTH * NUM_SAMPLES個取樣
    skip_second = 0.07
    swidth = 2
    '''
    #開啟聲音輸入
 
    save_buffer = []
    pa  =  pyaudio.PyAudio ()  
    stream  =  pa . open ( format = pyaudio.paInt16 ,  channels = 1 ,  rate = SAMPLING_RATE ,  input = True ,  #output = True,
                    frames_per_buffer = NUM_SAMPLES ) #,stream_callback = callback 
 
    time.sleep(0.1)
    stream.start_stream()
    skip_buffer = []
    num = 3500 / NUM_SAMPLES#SAMPLING_RATE / NUM_SAMPLES * skip_second
    for i in range(0, int(num)): 
        string_audio_data  =  stream.read ( NUM_SAMPLES) 
        skip_buffer.append(np.fromstring(string_audio_data, np.int16))
        #stream.read ( NUM_SAMPLES) 
    
    print('star recoder')
    
    num = SAMPLING_RATE / NUM_SAMPLES * recoder_second
    for i in range(0, int(num)): 
        #讀入NUM_SAMPLES個取樣
        string_audio_data  =  stream.read ( NUM_SAMPLES ) 
        save_buffer.append(np.fromstring(string_audio_data, np.int16))
        #stream.write(string_audio_data, NUM_SAMPLES)
    print("* done recording")
 
    stream.stop_stream()
    stream.close()
    pa.terminate()
    
    numpydata = np.hstack(save_buffer)
    skipnp = np.hstack(skip_buffer).tolist()
    
    '''
    #直接存檔
    WAVE_OUTPUT_FILENAME = filename + ".wav"
    wf = wave.open(WAVE_OUTPUT_FILENAME, 'wb')
    wf.setnchannels(1)
    wf.setsampwidth(pa.get_sample_size(pyaudio.paInt16))
    wf.setframerate(SAMPLING_RATE)
    wf.writeframes(b''.join(save_buffer))
    wf.close()
    '''
    if plotshow:
        if plotshow == 'line':
            plt.plot(numpydata,'r')
            plt.show()
        else:    
            plt.subplot(311)
            plt.plot(skipnp,'g')
            
            plt.subplot(312)
            plt.plot(save_buffer,'b')
            
            plt.subplot(313)
            plt.plot(numpydata,'r')
            plt.show()
    
    return numpydata, SAMPLING_RATE
 
if __name__ == '__main__':
    buffer = []
    recoder_second = 0.5
    buffer, fs = recoder(recoder_second, plotshow=True)
    fttmap, linef = fft_domyself(buffer, fs, plotshow = True)

開始錄音時發現會有一個雜訊，那個雜訊值非常大，以下是跳過雜訊再錄音0.05秒，green line是跳過的雜訊，blue line是接續綠色後面的0.05秒錄音

green line：跳過的雜訊
blue line：原始錄到的資料，每一點的資料量跟 (NUM_SAMPLES  =  100 # pyAudio內部緩存的塊的大小) 有關
red line：將blue line的資料串接起來