Flexible and efficient conversion from mono WAV to log-mel spectrogram

wav_to_mel.py

#!/usr/bin/env python
# -*- coding:utf-8 -*-


"""
"""


import librosa
import numpy as np


# ##############################################################################
# # AUDIO
# ##############################################################################
class WavToLogmel:
    """
    """
    def __init__(self, samplerate, winsize, hopsize, n_mels,
                 mel_fmin=50, mel_fmax=8_000, stft_window="hann"):
        """
        :param samplerate: Expected audio input samplerate.
        :param winsize: Window size for the STFT (and mel).
        :param winsize: Hop size for the STFT (and mel).
        :param stft_window: Windowing function for the STFT.
        :param n_mels: Number of mel bins.
        :param mel_fmin: Lowest mel bin, in Hz.
        :param mel_fmax: Highest mel bin, in Hz.
        """
        self.winsize = winsize
        self.hopsize = hopsize
        self.stft_window = stft_window
        self.mel_filt = librosa.filters.mel(sr=samplerate,
                                            n_fft=winsize, n_mels=n_mels,
                                            fmin=mel_fmin, fmax=mel_fmax)

    def __call__(self, wav_arr):
        """
        :param wav_arr: 1D audio array (float)
        :returns: log-mel spectrogram of shape ``(n_mels, t)``
        """
        stft_spec = np.abs(librosa.stft(y=wav_arr,
                                        n_fft=self.winsize,
                                        hop_length=self.hopsize,
                                        center=True,
                                        window=self.stft_window,
                                        pad_mode="reflect")) ** 2
        mel_spec = np.dot(self.mel_filt, stft_spec)
        logmel_spec = librosa.power_to_db(mel_spec, ref=1.0, amin=1e-10,
                                          top_db=None)
        return logmel_spec  # (nmels, t)