In [ ]:
from IPython.core.display import display, HTML
from IPython.display import Audio
from sklearn.neighbors import NearestNeighbors
import numpy as np
import pandas as pd
In [ ]:
def show_neighbors(index, n):
mfccs = ["mfcc_mean_%i"%idx for idx in range(1,13)] + ["mfcc_var_%i"%idx for idx in range(1,13)]
nbrs = NearestNeighbors(n_neighbors = n, algorithm='ball_tree').fit(df[mfccs].as_matrix())
example = df[mfccs].as_matrix()[index,:].reshape(1,-1) # reshape avoids a warning from sklearn wrt 1d arrays
distances, indices = nbrs.kneighbors(example)
for count, (distance, index) in enumerate(zip(distances[0], indices[0])):
print(' %i: %.2f %s (%s) %s' % (count + 1, \
distance,
df.loc[index, 'name'],
df.loc[index, 'path'], '(target sound)' if count == 0 else ''))
display(Audio(df.loc[index, 'path']))
In [ ]:
df = pd.DataFrame.from_csv('database.csv')
print(df.describe()) # Check statistics of descriptors
In [ ]:
df.query("spectral_centroid < 50") # Show the files with very low centroid
df.ix[101][["name","tags","path"]] # Select one
show_neighbors(101, 10)