Spaces:
Runtime error
Runtime error
amarchheda
commited on
Commit
•
a433359
1
Parent(s):
c1b97fa
added file
Browse files- main_code.py +80 -0
main_code.py
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import tensorflow as tf
|
3 |
+
from scipy.io.wavfile import write
|
4 |
+
import keras.backend as K
|
5 |
+
import librosa.display
|
6 |
+
import cv2
|
7 |
+
import librosa
|
8 |
+
import matplotlib.pyplot as plt
|
9 |
+
import librosa.display
|
10 |
+
import numpy as np
|
11 |
+
from keras.applications import VGG16
|
12 |
+
import os
|
13 |
+
import scipy
|
14 |
+
|
15 |
+
# Define function to preprocess input audio
|
16 |
+
#convert song to mel spectogram as siamese network doesn't work on sound directly
|
17 |
+
def create_spectrogram(clip,sample_rate,save_path):
|
18 |
+
plt.interactive(False)
|
19 |
+
fig=plt.figure(figsize=[0.72,0.72])
|
20 |
+
S=librosa.feature.melspectrogram(y=clip,sr=sample_rate)
|
21 |
+
librosa.display.specshow(librosa.power_to_db(S,ref=np.max))
|
22 |
+
fig.savefig(save_path,dpi=400,bbox_inches='tight',pad_inches=0)
|
23 |
+
plt.close()
|
24 |
+
fig.clf()
|
25 |
+
plt.close(fig)
|
26 |
+
plt.close('all')
|
27 |
+
del save_path,clip,sample_rate,fig,S
|
28 |
+
|
29 |
+
def load_img(path):
|
30 |
+
img=cv2.imread(path)
|
31 |
+
img=cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
|
32 |
+
img=cv2.resize(img,(150,150))
|
33 |
+
return img
|
34 |
+
|
35 |
+
import pickle
|
36 |
+
|
37 |
+
def main_loop():
|
38 |
+
|
39 |
+
with open('dict.pickle', 'rb') as handle:
|
40 |
+
songspecdict = pickle.load(handle)
|
41 |
+
|
42 |
+
# Load the song to match
|
43 |
+
song, sr = librosa.load("my_audio.wav")
|
44 |
+
to_match = np.copy(song[0:220500])
|
45 |
+
print("Loaded data into librosa...")
|
46 |
+
|
47 |
+
# Create spectrogram image of the song to match
|
48 |
+
create_spectrogram(to_match, sr, 'test.png')
|
49 |
+
print("Created spectogram...")
|
50 |
+
|
51 |
+
# Load the spectrogram image of the song to match
|
52 |
+
to_match_img = load_img('test.png')
|
53 |
+
to_match_img = np.expand_dims(to_match_img, axis=0)
|
54 |
+
print("Loaded spectrum image...")
|
55 |
+
|
56 |
+
# Get the embedding of the song to match
|
57 |
+
# Load the tune recognition model
|
58 |
+
model = tf.keras.models.load_model('./embdmodel_1.hdf5')
|
59 |
+
embedding_model=model.layers[2]
|
60 |
+
to_match_emb = embedding_model.predict(to_match_img)
|
61 |
+
print("Get song embedding...")
|
62 |
+
|
63 |
+
# Calculate the distances between the song to match and the songs in the database
|
64 |
+
songsdistdict = {}
|
65 |
+
for key, values in songspecdict.items():
|
66 |
+
dist_array = []
|
67 |
+
for embd in values:
|
68 |
+
dist_array.append(np.linalg.norm(to_match_emb - embd))
|
69 |
+
|
70 |
+
songsdistdict[key] = min(dist_array)
|
71 |
+
song_titles=list(songsdistdict.keys())
|
72 |
+
distances=list(songsdistdict.values())
|
73 |
+
|
74 |
+
# Get the title and artist of the recognized song
|
75 |
+
recognized_song_artist, recognized_song_title = song_titles[distances.index(min(distances))].split('-')
|
76 |
+
recognized_song_title = os.path.splitext(recognized_song_title)[0]
|
77 |
+
print(f'Artist: {recognized_song_artist}')
|
78 |
+
print(f'Title: {recognized_song_title}')
|
79 |
+
|
80 |
+
return recognized_song_title
|