wav2mel

Converts WAV audio [1] to Mel spectrograms for use in machine learning systems like Tacotron2.

This library contains portions of the copy-pasted code you see everywhere for WAV to Mel conversion.

[1] Or any audio format supported by librosa (which uses soundfile and audioread).

Installation

pip install wav2mel

Dependencies

• Python 3.6 or higher
• librosa, numpy, scipy, numba

Format

wav2mel outputs numpy save data: one .npy file each input file.

Usage

You can convert a single WAVE file from .wav to a mel spectrogram (.npy) as follows:

wav2mel < WAVE_FILE > NPY_FILE

Multiple WAVE files can also be converted and saved to a directory:

wav2mel --output-dir /path/to/mels WAVE_FILE [WAVE_FILE ...]

Each .npy file will be named after the corresponding .wav file.

See wav2mel --help for more options (filter/hop/window length, sample rate, etc.).

With GNU Parallel

find /path/to/wavs -name '*.wav' -type f | parallel -X wav2mel --output-dir /path/to/mels

mel2wav (Griffin-Lim)

You can convert a mel spectrogram to WAV audio too using griffin-lim:

mel2wav < NPY_FILE > WAVE_FILE

or

mel2wav --output-dir /path/to/wavs NPY_FILE [NPY_FILE ...]

See mel2wav --help for more options.