Skip to content

Instantly share code, notes, and snippets.

@hiromu
Created December 24, 2017 18:14
Show Gist options
  • Save hiromu/9909d9b6b88e66d2573deca43715b333 to your computer and use it in GitHub Desktop.
Save hiromu/9909d9b6b88e66d2573deca43715b333 to your computer and use it in GitHub Desktop.
Speaker Identification using GMM on MFCC
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import glob
import librosa
import numpy as np
import os
import sklearn.mixture
import sys
def load(audio_path):
y, sr = librosa.load(audio_path)
y_trim = librosa.effects.remix(y, intervals=librosa.effects.split(y))
mfcc = librosa.feature.mfcc(y=y_trim, sr=sr)
return mfcc.T
def fit(frames, test_ratio=0.05, n_components=16):
index = np.arange(len(frames))
np.random.shuffle(index)
train_idx = index[int(len(index) * test_ratio):]
test_idx = index[:int(len(index) * test_ratio)]
gmm = sklearn.mixture.GaussianMixture(n_components=n_components)
gmm.fit(frames[train_idx])
return gmm, frames[test_idx]
def predict(gmms, test_frame):
scores = []
for gmm_name, gmm in gmms.items():
scores.append((gmm_name, gmm.score(test_frame)))
return sorted(scores, key=lambda x: x[1], reverse=True)
def evaluate(gmms, test_frames):
correct = 0
for name in test_frames:
best_name, best_score = predict(gmms, test_frames[name])[0]
print 'Ground Truth: %s, Predicted: %s, Score: %f' % (name, best_name, best_score)
if name == best_name:
correct += 1
print 'Overall Accuracy: %f%%' % (float(correct) / len(test_frames))
if __name__ == '__main__':
gmms, test_frames = {}, {}
for filename in glob.glob(os.path.join(sys.argv[1], '*.wav')):
name = os.path.splitext(os.path.basename(filename))[0]
print 'Processing %s ...' % name
gmms[name], test_frames[name] = fit(load(filename))
evaluate(gmms, test_frames)
for filename in glob.glob(os.path.join(sys.argv[2], '*.wav')):
result = predict(gmms, load(filename))
print '%s: %s' % (os.path.basename(filename), ' / '.join(map(lambda x: '%s = %f' % x, result[:5])))
@Vichoko
Copy link

Vichoko commented Apr 18, 2022

You might try to train something simpler to identify issues in your code. For example, training with only 2 examples, and debugging the data as it navigates through the training module. Most errors are usually in between. After you manage to overfit that toy experiment, you should start increasing the data with such datasets.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment