On the efficiency of classical RASTA filtering for continuous speech recognition: Keeping the balance between acoustic pre-processing and acoustic modelling

Abstract

The efficiency of classical RASTA filtering for channel normalisation was investigated for continuous speech recognition based on context-independent and context-dependent hidden Markov models. For a medium and a large vocabulary continuous speech recognition task, recognition performance was established for classical RASTA filtering and compared to using no channel normalisation, cepstrum mean normalisation, and phase-corrected RASTA. Phase-corrected RASTA is a technique that consists of classical RASTA filtering followed by a phase correction operation. In this manner, channel bias is as effectively removed as with classical RASTA. However, for phase-corrected RASTA, amplitude drift towards zero in stationary signal portions is diminished compared to classical RASTA. The results show that application of classical RASTA filtering resulted in decreased recognition performance when compared to using no channel normalisation for all conditions studied, although the decrease appeared to be smaller for context-dependent models than for context-independent models. However, for all conditions, recognition performance was significantly and substantially improved when phase-corrected RASTA was used and reached the same performance level as obtained for cepstrum mean normalisation in some cases. It is concluded that classical RASTA filtering can only be effective for channel robustness, if the impact of the amplitude drift towards zero can be kept as limited as possible.

Introduction

The performance of automatic speech recognition (ASR) technology has been improving steadily, not in the least part thanks to improvements in acoustic feature extraction. Especially when the speech signal to be recognised is distorted, powerful pre-processing techniques are needed. In ASR over the telephone, one of the key factors introducing distortions is the communication channel.

The presence of a communication channel can, in principle, introduce four different types of distortions (e.g. Junqua and Haton, 1996). The channel may introduce:

• additive noise (e.g. telephone line clicks),

• linear filtering (e.g. the typical band-pass characteristic of 300–3400 Hz for fixed land-line telephony, and the linear frequency response of the handset microphone),

• non-linear filtering (as may be caused by using a carbon button microphone), and

• empty signal portions, which are typical for transmission problems in cellular telephony.

In this paper, we will restrict ourselves to the effects of linear filtering introduced by telephone channels. For this type of distortion it is the unpredictability of the linear transfer function that causes ASR performance to deteriorate.

Many different techniques have already been proposed to alleviate the effects of unpredictable linear filtering: cepstrum mean subtraction (Atal, 1974; Furui, 1981), using many different channels during training (e.g. Hirsch et al., 1991; Hermansky and Morgan, 1994; Aikawa et al., 1993; Nadeu et al., 1995; Junqua et al., 1995; de Veth and Boves, 1998), high-pass filtering of parameter tracks (Hirsch et al., 1991), the Gaussian dynamic cepstrum representation (Aikawa et al., 1993), RASTA filtering (Hermansky and Morgan, 1994), Slepian filtering (Nadeu et al., 1995), cepstral-time matrices (Milner and Vaseghi, 1995), phase-corrected RASTA filtering (de Veth and Boves, 1998), signal bias removal (Rahim and Juang, 1996), stochastic matching (Sankar and Lee, 1996), and combinations of these techniques (Junqua et al., 1995). An overview of many of these channel normalisation (CN) techniques can be found in (de Veth et al., 2001).

Given the rich diversity of different CN methods, the question arises which technique is best suited for a particular ASR design and a particular task. One of the factors to be considered is the type of units that are used to represent the speech sounds. In previous research on connected digit recognition, we used whole word models in addition to phone models (de Veth and Boves, 1998). For continuous speech recognition (CSR), however, there is no viable alternative other than some kind of sub-word models. The models can take various forms, yet in our research only continuous density mixture Gaussian hidden Markov models (HMMs) were used.

In this paper, we investigate the gain in recognition performance for four CN techniques in large vocabulary ASR systems that use either context-independent (CI) or context-dependent (CD) sub-word HMMs. The four CN techniques are (1) training on many different channels without additional processing (indicated as ‘no channel normalisation’ (NCN) in the remainder of the paper), (2) cepstral mean subtraction over the complete utterance (indicated as CMS), (3) classical RASTA filtering (indicated by clR), and (4) phase-corrected RASTA (de Veth and Boves, 1996, de Veth and Boves, 1997a, de Veth and Boves, 1997b, de Veth and Boves, 1998) (indicated by pcR). The aim of this paper is to study the interaction between the conventional RASTA filtering technique proposed for improvement of channel robustness on the one hand, and the average number of different left contexts per modelling unit on the other. The results for pcR enable identification of an important factor that limits the efficiency of clR. The results for NCN are included to indicate a baseline recognition performance. The results for CMS are included to indicate the level of what ideally can be achieved when a filtering approach to CN is chosen. So, the results for NCN and CMS will serve to indicate a background against which the results obtained with clR and pcR can be more readily interpreted.

This paper is further organised as follows. In Section 2, the salient features of clR and pcR are discussed, and our research questions are precisely formulated. The telephone speech databases that were used in our experiments are described in Section 3. In Section 4, the signal processing for our experiments is described. The topology of the HMMs, the way these were trained and the recognition task are described in Section 5. In addition, a measure for the average number of left contexts per model is introduced in Section 5. The results of our recognition experiments are presented in Section 6 and discussed in Section 7. Finally, in Section 8, the main conclusions are summarised.