awjuliani / sound-cnn Goto Github PK

Hello guys, im trying to run this project and i changed the operator / to // since im using python 3. I'm still having errors:

(pythonProject3) C:\Users\User\PycharmProjects\pythonProject3>python C:\Users\User\Desktop\adsp\train.py 240 44100 C:\Users\User\Desktop\audio\ 1000 1
Traceback (most recent call last):
File "C:\Users\User\Desktop\adsp\train.py", line 19, in
classes, trainX, trainYa, valX, valY, testX, testY = util.processAudio(bpm, samplingRate, mypath)
File "C:\Users\User\Desktop\adsp\utilities.py", line 26, in processAudio
audData = np.reshape(audData[1][:, 1][0:samples * ((seconds * samplingRate) // samples)],
ZeroDivisionError: integer division or modulo by zero

By using spectrogram, we can simply treat audio recognition as image recognition, so all the technologies used at image recognition can be used at audio recognition as well

This example uses the same samples as speech_commands, that is wave audio files of people saying ten different words('yes', 'no', 'up', 'down', 'left', 'right', 'on', 'off', 'stop', 'go') from Speech Commands dataset

First, transform the original audio files into spectrograms, we use the following python code to do the task:

import numpy as np
from scipy.io import wavfile

X_SIZE = 16000
IMG_SIZE = 128

def spectrogram(filepath):
	framerate, wav_data = wavfile.read(filepath)

	window_length = 512
	window_shift = 121

	if len(wav_data) > X_SIZE:
		wav_data = wav_data[:X_SIZE]

	X = np.zeros(X_SIZE).astype('float32')
	X[:len(wav_data)] += wav_data
	spec = np.zeros((IMG_SIZE, IMG_SIZE)).astype('float32')

	for i in range(IMG_SIZE):
		start = i * window_shift
		end = start + window_length
		sig = np.abs(np.fft.rfft(X[start:end] * np.hanning(window_length)))
		spec[:,i] = (sig[1:IMG_SIZE + 1])[::-1]

	spec = (spec-spec.min())/(spec.max()-spec.min())
	spec = np.log10((spec * 100 + 0.01))
	spec = (spec-spec.min())/(spec.max()-spec.min()) - 0.5

	return spec

For example, the spectrogram for test.wav is:

To reduce the amount of computation, we limit the size of spectrogram to 128×128, so the model can be defined as follow:

Run this model, you will get a test accuracy around 93%(while train accuracy is 98%) after 15 epochs

Reference

• Simple Audio Recognition

Github Link

student@student-Vostro-3669:~/sound-cnn$ python train.py 240 44100 /home/student/traindata 1000 150
Traceback (most recent call last):
File "train.py", line 20, in
classes,trainX,trainY,valX,valY,testX,testY = util.processAudio(bpm,samplingRate,mypath)
File "/home/student/sound-cnn/utilities.py", line 29, in processAudio
Ys = np.concatenate(labelList)
ValueError: need at least one array to concatenate

Hi. First, I tested mono .wav files which resulted in an error.

Now I'm using stereo .wav files which results in another error.

stefan@stefan-quadcore ~/dev/sound-cnn-master $ ./doit.sh 
Traceback (most recent call last):
  File "train.py", line 20, in <module>
    classes,trainX,trainYa,valX,valY,testX,testY = util.processAudio(bpm,samplingRate,mypath)
  File "/home/stefan/dev/sound-cnn-master/utilities.py", line 24, in processAudio
    audData = np.reshape(audData[1][:,1][0:samples*((seconds*samplingRate)/samples)],[samples,(seconds*samplingRate)/samples])
TypeError: slice indices must be integers or None or have an __index__ method
stefan@stefan-quadcore ~/dev/sound-cnn-master $ cat doit.sh 
python3 train.py 240 44100 ../sounds-stereo/ 1000 100

Why does stuff not work in this world?

Cheers

How to use use this model to make new predictions using a new audio file ?
i've tried to reload the model.ckpt file but it keeps giving me errors ...

python train.py 240 44100 audio/ 1000 150

Traceback (most recent call last):
File "train.py", line 20, in
classes,trainX,trainYa,valX,valY,testX,testY = util.processAudio(bpm,samplingRate,mypath)
File "/home/fuquo/src/sound-cnn/utilities.py", line 22, in processAudio
seconds = audData[1][:,1].shape[0]/samplingRate
IndexError: too many indices for array

Could you setup a quick demo?
I cannot understand why any input file to be theorized as a new class and i am getting error on train.py:
'ValueError: Cannot feed value of shape (1, 2) for Tensor u'Placeholder_1:0', which has shape '(?, 4)'