Over the past several years there has been considerable attention focused on the problem of enhancement and bandwidth compression of speech degraded by additive background noise. This interest is motivated by several factors including a broad set of important applications, the apparent lack of robustness in current speech-compression systems and the development of several potentially promising and practical solutions. One objective of this paper is to provide an overview of the variety of techniques that have been proposed for enhancement and bandwidth compression of speech degraded by additive background noise. A second objective is to suggest a unifying framework in terms of which the relationships between these systems is more visible and which hopefully provides a structure which will suggest fruitful directions for further research.

Enhancement and bandwidth compression of noisy speech

A technique for robust spectral representation of all-pole sequences is proposed. It is shown that the autocorrelation of an all-pole sequence, obtained by passing white noise through an all-pole filter 1/A(z), is an all-pole sequence of the form 1/A/sup 2/(z). A short-time modified coherence (SMC) representation is proposed that is an all-pole modeling of the autocorrelation sequence with a spectral shaper. The spectral shaper, essentially a square root operator in the frequency domain, compensates for the inherent spectral distortion introduced by the autocorrelation operation on the autocorrelation sequence of the signal. The properties of the SMC representation, especially its robustness to additive white noise, are analyzed. Initial implementation of the SMC in a speaker-dependent isolated word recognizer shows an improvement in recognition accuracy equivalent to an increase in input SNR of approximately 13 dB, as compared to the LPC recognizer. >

The short-time modified coherence representation and noisy speech recognition

A new idea, enhancing speech based on auditory evidence, is explored for the problem of enhancing speech degraded by stationary and nonstationary additive white noise. Distinguishing different objectives for heavy and light noise interference, two related algorithms are developed. For speech degraded by heavy noise, the improvement in signal-to-noise ratio (SNR) is as high as 12 dB; for lightly noisy speech, the improvement is modest and decreases as the SNR of the noisy speech increases. Quantizing noise is used to assess the capacity for reducing nonstationary noise using these algorithms; a significant reduction of such noise and an improvement in speech quality are achieved. The advantages of the proposed algorithms for speech enhancement include no need for prior knowledge of the noise and only a modest computational requirement. >

Speech enhancement based conceptually on auditory evidence

A technique for robust spectral representation of all-pole sequences is proposed. It is shown that the autocorrelation of an all-pole sequence, obtained by passing white noise through an all-pole filter 1/A(z), is an all-pole sequence of the form 1/A/sup 2/(z). The short-time modified coherence (SCM) representation, proposed here, is an all-pole modeling of the autocorrelation sequence followed by a spectral shaper. The spectral shaper, essentially a square root operator in the frequency domain, compensates for the inherent spectral distortion introduced by the autocorrelation operation on the autocorrelation sequence of the signal. The properties of the SMC representation, especially its robustness to additive white noise, are analyzed. Initial implementation of the SMC in a speaker-dependent isolated-word recognizer shows a considerable improvement over the standard linear predictive coding (LPC) representation. The SMC recognizer achieved an improvement in recognition accuracy equivalent to an increase in input SNR of approximately 13 dB, as compared to the LPC recognizer. >

The short-time modified coherence representation and its application for noisy speech recognition

Processing of the speech signal in the autocorrelation domain in the context of robust feature extraction is based on the following two properties: 1) pole preserving property (the poles of a given (original) signal are preserved in its autocorrelation function), and 2) noise separation property (the autocorrelation function of a noise signal is confined to lower lags, while the speech signal contribution is spread over all the lags in the autocorrelation function, thus providing a way to eliminate noise by discarding lower-lag autocorrelation coefficients). In this paper, we use these properties to derive robust features for automatic speech recognition. We compute the magnitude spectrum of the one-sided higher-lag autocorrelation sequence, process it through a Mel filter bank and parameterise it in terms of Mel Frequency Cepstral Coefficients (MFCCs). Since the proposed method combines autocorrelation domain processing with Mel filter bank analysis, we call the resulting MFCCs, Autocorrelation Mel Frequency Cepstral Coefficients (AMFCCs). Recognition experiments are conducted on the Aurora II database and it is found that the AMFCC representation performs as well as the MFCC representation in clean conditions and provides more robust performance in the presence of background noise.

MFCC computation from magnitude spectrum of higher lag autocorrelation coefficients for robust speech recognition

A speech processing system named SPAC (SPlicing of AutoCorrelation function) is proposed in order to compress or expand the speech spectrum, to prolong or shorten the duration of utterance, and to reduce the noise level in speech signal. A period of short-time autocorrelation function is sampled and spliced after change of the time scale. Transformed speech is quite natural and free from distortion. Applications of SPAC are expected in many fields such as improvement of speech quality, narrow band transmission, communication aid for hard of hearing, information service for blind, unscrambling of helium speech, stenography and so on.

J. Suzuki

Papers

Speech processing by splicing of autocorrelation function