Analysis of CNN-based speech recognition system using raw speech as input

Abstract Automaticspeechrecognitionsystemstypicallymodeltherela-tionship between the acoustic speech signal and the phones intwo separate steps: feature extraction and classier training. Inourrecentworks, wehaveshownthat, intheframeworkofcon-volutionalneuralnetworks(CNN),therelationshipbetweentheraw speech signal and the phones can be directly modeled andASR systems competitive to standard approach can be built. Inthis paper, we rst analyze and show that, between the rst twoconvolutional layers, the CNN learns (in parts) and models thephone-specic spectral envelope information of 2-4 ms speech.Given that we show that the CNN-based approach yields ASRtrends similar to standard short-term spectral based ASR sys-tem under mismatched (noisy) conditions, with the CNN-basedapproach being more robust.Index Terms: automatic speech recognition, convolutionalneural networks, raw signal, robust speech recognition. 1. Introduction State-of-the-art automatic speech recognition (ASR) systemstypically model the relationship between the acoustic speechsignal and the phones in two separate steps, which are op-timized in an independent manner [1]. In a rst step, thespeech signal is transformed into features, usually composed ofa dimensionality reduction phase and an information selectionphase, based on the task-specic knowledge of the phenomena.These two phases have been carefully hand-crafted, leading tostate-of-the-art features such as Mel frequency cepstral coef-cients(MFCCs)orperceptuallinearpredictioncepstralfeatures(PLPs). In a second step, the likelihood of subword units suchas, phonemes is estimated using generative models or discrimi-native models.In recent years, in the hybrid HMM/ANN framework [1],there has been growing interests in using “intermediate” rep-resentations instead of conventional features, such as cepstral-based features, as input for neural networks-based systems.ANNs with deep learning architectures, more precisely, deepneural networks (DNNs) [2, 3], which can yield better systemthan a single hidden layer MLP have been proposed to addressvarious aspects of acoustic modeling. More specically, useof context-dependent phonemes [4, 5]; use of spectral featuresas opposed to cepstral features [6, 7]; CNN-based system withMel lter bank energies as input [8, 9, 10]; combination of dif-ferent features [11], to name a few. Features learning from therawspeechsignalusingneuralnetworks-basedsystemshasalsobeen investigated in [12]. In all these approaches, the features