Resampling to the same sample rate via Python

Question

I'm working on machine learning model and tried to downsample WAV files with Python script, but it didn't work, so I made the simple version - script "resamples" WAV file from 44100Hz to the same 44100Hz sample rate. Original and output files seems identical while listening to, they are both 16-bit depth, 44100Hz sample rate, output has 1 channel, but I did it for testing purposes and I think shouldn't matter.

However, plot shows minimal differences original vs "resampled" signal:

There is my whole code:

def calculate_differences_between_arrays(arr1, arr2):
    differences = []
    for val1, val2 in zip(arr1, arr2):
        diff = val1 - val2
        differences.append(diff)
    return differences

def show_plot_for_list(list, header):
    plt.plot(list)
    plt.title(header)
    plt.show()

if __name__ == "__main__":

    # audio read
    origin_sample_rate, origin_audio = wavfile.read('org.wav')
    wavfile.write('out_org.wav', origin_sample_rate, origin_audio[:,0])

    origin_num_samples, origin_num_channels = origin_audio.shape

    # resampling
    target_audio_scipy = scipy.signal.resample(origin_audio[:,0], origin_num_samples).astype(int)
    target_audio_scipy = np.array(target_audio_scipy, np.int16)

    # show differences
    diffArray = calculate_differences_between_arrays(origin_audio[:,0], target_audio_scipy)
    show_plot_for_list(diffArray, 'Differences')

    # write modified audio to WAV file
    wavfile.write('out.wav', origin_sample_rate, target_audio_scipy)

Is my script wrong, or it is normal behavior and it won't be problem while downsampling for example to 16000Hz?

Answer 1

There is no problem with the resample command of sampling to the same sampling rate, and the 1 count error is simply due to rounding with the 16 bit precision used. Zooming in on an example audio track I confirmed the OP's result of having the input and output matching within 1 sample due to rounding and not a time offset.

Extreme zoom-in:

Resampling to 16 KHz will result in larger deviations as expected by reducing all spectrum above ~ 7 KHz and indicated in the comparative plot below. The audio should be completely recognizable however:

After resampling to 16 KHz

My code with resampling done:

# audio read
origin_sample_rate, origin_audio = wavfile.read('./data/Inspiring-Powerful-Cinematic_AdobeStock_462873176.wav')
wavfile.write('./out_org.wav', origin_sample_rate, origin_audio[:,0])

origin_num_samples, origin_num_channels = origin_audio.shape

new_samps = int(origin_num_samples * 16000/44100)

# resampling
target_audio_scipy = sig.resample(origin_audio[:,0], new_samps).astype(int)
target_audio_scipy = np.array(target_audio_scipy, np.int16)