Cepstral compensation by polynomial approximation for environment-independent speech recognition

Abstract

Speech recognition systems perform poorly on speech degraded by even simple effects such as linear filtering and additive noise. One possible solution to this problem is to modify the probability density function (PDF) of clean speech to account for the effects of the degradation. However, even for the case of linear filtering and additive noise, it is extremely difficult to do this analytically. Previously attempted analytical solutions to the problem of noisy speech recognition have either used an overly-simplified mathematical description of the effects of noise on the statistics of speech, or they have relied on the availability of large environmentspecific adaptation sets. Some of the previous methods required the use of adaptation data that consists of simultaneously-recorded or "stereo" recordings of clean and degraded speech. In this paper we introduce an approximation-based method to compute the effects of the environment on the parameters of the PDF of clean speech.In this work, we perform compensation by Vector Polynomial approximationS (VPS) for the effects of linear filtering and additive noise on the clean speech. We also estimate the parameters of the environment, namely the noise and the channel, by using piecewiselinear approximations of these effects.We evaluate the performance of this method (VPS) using the CMU SPHINX-II system and the 100-word alphanumeric CENSUS database. Performance is evaluated at several SNRs, with artificial white Gaussian noise added to the database. VPS provides improvements of up to 15 percent in relative recognition accuracy.

Uncertainty in Signal Estimation and Stochastic Weighted Viterbi Algorithm: A Unified Framework to Address Robustness in Speech Recognition and Speaker Verification

“…This assumption is very important because it dramatically reduces the number of parameters to estimate and the amount of adaptation data required. Despite the fact that (27) to estimate d M is the same expression employed to estimate convolutional distortion (Acero & Stern, 1990) if additive noise is not present (Yoma, 1998-B), the methods in (Acero & Stern, 1990;Moreno et al, 1995;Raj et al, 1996) do not compensate the HMMs. Notice that the effect of the transfer function that represents a linear channel is supposed to be an additive constant in the log-cepstral domain.…”

“…al, 2006) proposes a model of the low bit rate coding-decoding distortion that is different from the model of the additive and convolutional noise, although they are similar to some extent. The mean and variance compensation is code-word dependent in (Acero & Stern, 1990;Moreno et al, 1995;Raj et al, 1996). In contrast, V are considered independent of the code-word in (Yoma et.…”

“…When compared to MLLR (Gales, 1998), the proposed computation of the coding-decoding distortion requires fewer parameters to estimate, although it should still lead to high improvements in word accuracy as a speaker adaptation method. Finally, the method discussed in this section to estimate the coding-decoding distortion is similar to (Acero and Stern, 1990;Moreno et al, 1995;Raj et al, 1996) to compensate additive/convolutional noise and estimate the unobserved clean signal. In those papers the p.d.f.…”

Acoustic Feature Optimization Based on F-Ratio for Robust Speech Recognition

“…Speaker robust technologies comprise speaker normalization and model adaptation, such as speaker-adaptive training (SAT) [3], vocal tract length normalization (VTLN) [4]. Noise robustness and compensation technologies comprise the parallel model combination (PMC) [5], Vector Taylor Series (VTS/VPS) [6], [7], Quantile based Histogram Equalization (HEQ) [8], Switching Linear Dynamic Model (SLDM) [9], and etc. Other types of variations also attract more and more attentions, such as the emotion affected speech recognition (EASR).…”

On the Effects of Filterbank Design and Energy Computation on Robust Speech Recognition

“…Other related work includes speech enhancement [4], normalization of the noisy features statistical properties [5]- [7], and dynamic feature combinations [8]. In [9]- [11], the effect of environmental noise on the statistical speech models was investigated and two algorithms (CDCN and MFCDCN) were proposed for compensating it. However, the feature robustness problem remains unsolved in a globally optimal way.…”