Spectral entropy based feature for robust ASR

Abstract

Abstract. In general, entropy gives us a measure of the number of bits required to represent some information. When applied to probability mass function (PMF), entropy can also be used to measure the "peakiness" of a distribution. In this paper, we propose using the entropy of short time Fourier transform spectrum, normalised as PMF, as an additional feature for automatic speech recognition (ASR). It is indeed expected that a peaky spectrum, representation of clear formant structure in the case of voiced sounds, will have low entropy, while a flatter spectrum corresponding to non-speech or noisy regions will have higher entropy. Extending this reasoning further, we introduce the idea of multi-band/multi-resolution entropy feature where we divide the spectrum into equal size sub-bands and compute entropy in each sub-band. The results presented in this paper show that multi-band entropy features used in conjunction with normal cepstral features improve the performance of ASR system. 2 IDIAP-RR 03-56

Multi-resolution Spectral Entropy Feature for Robust ASR

“…Instead of transforming the spectral information into cepstral domain, we suggested computing entropy from the sub-bands of spectrum and trying to locate the spectral peaks of the spectrum which are supposed to be more robust to noise. In [4], we showed that the proposed multiresolution spectral entropy feature is not very competitive when compared to the state-of-the-art PLP cepstral features, but improves the robustness of the ASR system when appended to the PLP features.…”

“…The importance of formants is well know and in [7] the author has tried to use the location of spectral peaks as an additional feature in ASR. On the similar lines, the central idea in [4] while using multi-resolution spectral entropy as a feature was to capture the peaks of the spectrum and their location. To compute entropy of a spectrum we converted the spectrum into a PMF like function by normalizing it.…”

“…Realizing that full-band entropy can capture only the gross peakiness of the spectrum but not the location of the formants, in [4] we suggested the idea of multi-resolution/multi-band entropy feature to capture the location of the formants. To extract multi-band entropy features, we divided the full-band spectrum into J non-overlapping sub-bands of equal size.…”

“…In [4] we advised that the way full-band spectrum is converted into a PMF, each sub-band spectrum should be converted into a sub-band PMF and entropy be computed for each sub-band PMF. Converting each sub-band into a PMF enhances the false local peaks as the normalization was done only for that sub-band.…”

“…We followed a completely different approach while proposing our multi-resolution spectral entropy feature [4]. Instead of transforming the spectral information into cepstral domain, we suggested computing entropy from the sub-bands of spectrum and trying to locate the spectral peaks of the spectrum which are supposed to be more robust to noise.…”