Robust speech recognition using features based on zero crossings with peak amplitudes

Abstract

This paper presents an extensive study of zero crossings with peak amplitudes (ZCPA) features, that have earlier been shown to outperform both conventional and auditory-based features in presence of additive noise. The study starts by optimizing different parameters involved in ZCPA feature computation, followed by a comparison of ZCPA and MFCC features on two recognition tasks in different background conditions. The main differences between the two feature types were identified, and their individual effects on ASR performance were evaluated. The importance of a proper choice of analysis frame lengths and filter bandwidths in ZCPA feature extraction was demonstrated. Furthermore, the use of dominant frequency information in ZCPA features was found to be a major reason for increased robustness of ZCPA features compared to MFCC features.

Perceptual features for automatic speech recognition in noisy environments

“…These may be used for applications such as in human-computer interactions, document preparation and dictation, telephone voice response systems, database access, hands-free applications as in car phones or voice-enabled PDAs, web enabling via voice, to name but a few. In spite of focused research in this field for the past several decades, the understanding of the acousticphonetic characteristics of speech, speech variability and speech perception is far from complete, and robust speech recognition with high reliability has not been achieved (Kim et al, 1999;Abdelatty et al, 1999;Gajić and Paliwal, 2003). The speech recognition process may work well in clean conditions but degrades significantly in speaker and channel mismatch conditions.…”

Auditory Features Revisited for Robust Speech Recognition

“…The ZCPA model has many variable parameters. The scope of this paper covers a limited number of optimisations (using Gajic's recommended values [2] as a starting point). However there are many further potential adaptations possible which may bring about increased performance -frame length, delta window, filterbank parameters, histogram parameters, number of coefficients etc.…”

“…Subsequent implementations [2,1] use FIR filterbanks exclusively. A filterbank was designed with 16 Hamming FIR filters of order 61.…”

“…The centre frequencies of the filters were spread evenly on the Bark scale from 200Hz to 3400Hz (see Figure 4). Filter bandwidths were equal to 2 bark -reported to be an optimal spacing by Gajic [2].…”

Asymptotic theory for curve-crossing analysis

“…In particular, C n (0) is the zero-crossing count. Level-crossing analysis has been widely used in engineering, physics, speech recognition and other fields; see [11,[19][20][21]30,35]. For example, if X i = k j=1 [A j cos(iω j ) + B j sin(iω j )] + e i for independent normal random variables A j , B j and e i , then one can estimate the frequencies ω j by considering the zero-crossing counts of X i and its differencing sequences; see [21].…”