tflibrosa

re-implementation of torch librosa for tensorflow. It is usefull if you want to compute Spectrogram on GPU for faster inference instead of using librosa.

Installation

pip install tflibrosa

Example

To do some inference on single sample, you can use python script in examples/ folder or use as follows:

import numpy as np 
from tflibrosa import STFT, Spectrogram, LogmelFilterBank
import librosa
import tensorflow as tf 
audio = np.random.uniform(0,1 ,(32000 * 5))
print(audio.shape)

sample_rate = 32000
n_fft = 2048
hop_size = 512
window = 'hann'
pad_mode = 'reflect'
mel_bins = 64
ref = 1.0
amin = 1e-10
fmin = 20
fmax = 16000 
top_db = 80.0
center = True 
dtype=None

spectrogram_extractor = Spectrogram(n_fft=n_fft, hop_length=hop_size, 
                win_length=n_fft, window=window, center=center, pad_mode=pad_mode, 
                freeze_parameters=True, dtype="float32")

# Logmel feature extractor
logmel_extractor = LogmelFilterBank(sr=sample_rate, n_fft=n_fft, is_log=True, 
    n_mels=mel_bins, fmin=fmin, fmax=fmax, ref=ref, amin=amin, top_db=top_db, 
    freeze_parameters=True, dtype="float32")


spectrogram = spectrogram_extractor(audio[None, :])

mel_spectrogram = logmel_extractor(spectrogram)

print(mel_spectrogram) # (batch size, num_channels, timestamps)

Acknowledgement

• librosa : https://librosa.org/doc/latest/index.html
• torchlibrosa : https://github.com/qiuqiangkong/torchlibrosa