Learning linearly separable features for speech recognition using convolutional neural networks

TL;DR

This paper explores using convolutional neural networks with a linear classifier to learn linearly separable features directly from raw speech signals, achieving comparable or better speech recognition performance than traditional cepstral features.

Contribution

It introduces a CNN-based approach with a linear classifier that learns directly from raw speech, simplifying the feature extraction process and improving recognition performance.

Findings

Linear CNN features match or outperform cepstral features.

Simplifies speech recognition pipeline by learning features directly from raw data.

Achieves comparable or better results with fewer parameters.

Abstract

Automatic speech recognition systems usually rely on spectral-based features, such as MFCC of PLP. These features are extracted based on prior knowledge such as, speech perception or/and speech production. Recently, convolutional neural networks have been shown to be able to estimate phoneme conditional probabilities in a completely data-driven manner, i.e. using directly temporal raw speech signal as input. This system was shown to yield similar or better performance than HMM/ANN based system on phoneme recognition task and on large scale continuous speech recognition task, using less parameters. Motivated by these studies, we investigate the use of simple linear classifier in the CNN-based framework. Thus, the network learns linearly separable features from raw speech. We show that such system yields similar or better performance than MLP based system using cepstral-based features as input.