Reduced Channel Dependence for Speech Recognition

REDUCED CHANNEL DEPENDENCE FOR SPEECH RECOGNITION Hy Murveit, John Butzberger, and Mitch Weintraub SRI International Speech Research and Technology Program Menlo Park, CA, 94025 1. ABSTRACT Speech recognition systems tend to be sensitive to unimportant steady-state variation in speech spectra (i.e. those caused by vary- ing the microphone or channel characteristics). There have been many attempts to solve this problem; however, these techniques are often computationally burdensome, especially for real-time implementation. Recently, Hermansy et el. [1] and Hirsch et el. [2] have suggested a simple technique that removes slow-moving lin- ear channel variation with little adverse effect on speech recogni- tion performance. In this paper we examine this technique, known as RASTA filtering, and evaluate its performance when applied to SRI's DECIPHER TM speech recognition system [3l. We show that RASTA filtering succeeds in reducing DECIPHERTM's depen- dence on the channel. 2. INTRODUCTION A number of techniques have been developed to compen- sate for the effects that varying microphone and channels have on the acoustic signal. Erell and Weintraub [4, 5] have used additive corrections in the filter-bank log energy or cepstral domains based on equalizing the long-tenn average of the observed filter-bank log energy or cepstral vector to that of the training data. The techniques developed by Rose and Paul [6] and Acero [7] used an iterative technique for estimating the cepstral bias vector that will maximize the likelihood of the input utterance. Nadas et el. [8] used an adaptive linear transformation applied to the input repre- sentation, where the adaptation uses the VQ distortion vec- tor with respect to a predefined codebook. VanCompemolle [10] scaled the filter-bank log energies to a specified range using running histograms, and Rohlicek [9] experimented with a number of histogram-based compensation metrics based on equalizing different aspects of the probability dis- tribution. One important limitation of the above approaches is that they rely on a speech/nonspeech detector. Each of the above approaches computes spectral properties of the input speech sentence and subsequently compensates for the sta- tistical differences with certain properties of the training data. If the input acoustical signal is not segmented by sen- tence (e.g. open microphone with no push-to-talk button) and there are long periods of silence, the above approaches would not be able to operate without some type of reliable automatic speech-input/sentence-detection mechanism. An automatic sentence-detection mechanism would have con- siderable difficulty in reliably computing the average speech spectrum if there were many other nonspeech sounds m the environment. A second class of techniques developed around auditory models (Lyon [11]; Cohen [12]; Seneff [13]; Ghitza [14]). These techniques use various automatic gain control and other auditory-type modeling techniques to output a spec- tral vector that has been adapted based on the acoustic his- tory. A potential limitation of this approach is that many of these techniques are very computationally intensive. 3. THE RASTA FILTER RASTA filtering is a high-pass filter applied to a log-spec- tral representation of speech. It removes slow-moving vari- ations from the log spectrum. The filtering …

Content truncated to 500 words