I'm trying to define a function which:

1. Reads in each 1min audio file from a directory
2. Calculates features for each second of each 1min file, returning the numpy array feats of shape (60, 96, 64) for each file, where 60 denotes each second
3. Takes the mean across all seconds in a 1min file to return the array of features_from_one_file shape (96, 64)
4. Appends each of these mean arrays to the 3D array features_allfiles, so that each 1min file is represented as a dimension (correct term?) in features_allfiles. e.g if five 1min files were used, this would have shape (5, 96, 64)
5. I then aim to adapt this so that any files n minutes in length will have their feats arrays split by n, so that the average feats are returned on a per minute basis.

I've got stuck at step four, so could use help with this, any suggestions for step 5 also welcomed!

Here's my code so far:

def get_features(directory):
audio_fs = os.listdir(directory) #list of all files in directory

features_allfiles = np.empty([0,96,64])

for f in audio_fs:
    #find file:
    path = os.path.join(directory, f)
    
    #calculate features from audio file:
    feats = vggish_input.wavfile_to_examples(path)
    print(np.shape(feats))                    #this returns (62, 96, 64) for a 1min file
    
    #Get the mean of the these 62 2D arrays
    features_from_one_file = np.mean(feats, axis = 0) 
    print(np.shape(features_from_one_file))   #this returns (96, 64)
    
    #Append the mean of each file to features_allfiles, so that it has shape (n, 96, 62), where n = number of files
    ???
    
return features_allfiles

CodePudding user response：

You can use np.vstack but first you have to add a new dimension to features_from_one_file:

EDIT

As suggested @hpaulj, the best practice is to stack only once, so you need to collect all `features_from_one_file` in a list and then stack them:

import numpy as np

all_feats = []
for i in range(5):
    #create features    
    features_from_one_file = np.random.randn(96,64)
    #append Broadcasting first to have dimension 1x96x64
    all_feats.append(features_from_one_file[None,:] )

#stack all at once
features_allfiles = np.vstack(all_feats)

print(features_allfiles.shape)

wich outputs

(5, 96, 64)

Old propossal: stack by one (this is slower)

features_allfiles = np.empty([0,96,64])

for i in range(5):
    #new features
    features_from_one_file = np.random.randn(96,64)
    
    #vertical stack, [None,:] recast array adding new dimension first
    # you can also use features_from_one_file.reshape(1,96,64)
    features_allfiles = np.vstack([features_allfiles,
                                   features_from_one_file[None,:]])


print(features_allfiles.shape)

wich outputs

(5, 96, 64)