gorillabus / urban-audio-classifier Goto Github PK

why didnt u use 10cross valitation?
May u show that in a note book

Running 1-data-analysis.ipynb gives an error in line

helpers.play_dataset_sample(row, audio_path)

because include/helpers.py is missing
import IPython as IP

Hi,
Thank you for sharing your code. By the way, I found that in your code for get_mel_spectrogram, there is something wrong with changing a Mel spectrogram on a decibel scale.

# Generates/extracts Log-MEL Spectrogram coefficients with LibRosa 
def get_mel_spectrogram(file_path, mfcc_max_padding=0, n_fft=2048, hop_length=512, n_mels=128):
    try:
        # Load audio file
        y, sr = librosa.load(file_path)

        # Normalize audio data between -1 and 1
        normalized_y = librosa.util.normalize(y)

        # Generate mel scaled filterbanks
        mel = librosa.feature.melspectrogram(normalized_y, sr=sr, n_mels=n_mels)

        # Convert sound intensity to log amplitude:
        mel_db = librosa.amplitude_to_db(abs(mel))

        # Normalize between -1 and 1
        normalized_mel = librosa.util.normalize(mel_db)

     ... 
    return normalized_mel

Here you are using librosa.amplitude_to_db but you should use librosa.power_to_db for Mel spectrograms. Because librosa.melspectrogram returns only a power spectrogram (amplitude squared) unless you set its parameter as power=1.

You can find details from here: https://librosa.org/doc/main/generated/librosa.feature.melspectrogram.html
Please see its examples.

And its source code here: https://librosa.org/doc/main/_modules/librosa/feature/spectral.html