Spafe: Simplified python audio featuresextraction
Ayoub Malek
Abstract
This paper describes version 0.3.1 of spafe: a python package for audio features extraction based on the Numpy (Harris et al., 2020) and Scipy (Pauli Virtanen & Vázquez-Baeza, 2019) libraries.Spafe implements various features extraction techniques that can be used to solve a wide variety of recognition and classification tasks (speaker verification, spoken emotion recognition, spoken language identification etc.).The paper provides a brief overview of the library's structure, theory and functionalities. Statement of needIn speech processing, features extraction is essentially the estimation of a parametric representation of an input signal.This is a key step in any audio based modeling and recognition process (e.g.speech recognition, sound classification, speaker authentication etc.).There are several speech features to extract, such as the Linear Frequency Cepstral Coefficients (LFCC), Mel Frequency Cepstral Coefficients (MFCC), Linear Predictive Coding (LPC), and Constant-Q Cepstral Coefficients (CQCC) etc.Each type of features has its own advantages and drawbacks (e.g.noise robustness, complexity, inter-components correlation etc.) that can directly affect the researched topic.Unfortunately, existing libraries for extracting these features (e.g.librosa (McFee et al., 2015), python_speech_features (Lyons et al., 2020), SpeechPy (Torfi, 2018) and Bob (A. Anjos & Marcel, 2017)) are limited and mostly focus on one extraction technique (e.g.MFCC), thus it is hard to find reliable implementations of other features extraction algorithms.Consequently, this slows down the research and hinders the possibility of exploring and comparing these different approaches.Hence, the need for spafe, a straightforward solution that unites all these different techniques in one python package.