Speech analysis and recognition using interval statistics generated from a composite auditory model

Abstract

A modeling approach to auditory speech analysis and recognition is proposed and evaluated, where a composite auditory model is used to generate parallel sets of auditory-nerve instantaneous firing rates (IFR's) along the spatial dimension, followed by a processing stage that constructs from the IFR's an interval statistics in a form called the interpeak interval histogram (IPIH). A speech preprocessor is designed that performs transformation on the auditory IPIH's and interfaces the IPIH-based auditory representation with a hidden Markov model-based (HMM-based) speech recognizer. The results demonstrate that the new preprocessor consistently outperforms the conventional mel frequency cepstral coefficient-based (MFCC-based) preprocessor for the signal-tonoise ratio (SNR) level up to at least 16 dB.

The robustness of speech representations obtained from simulated auditory nerve fibers under different noise conditions

“…Our study confirms previous reports that timing features derived from an auditory model give a substantial performance benefit compared to MFCCs in Gaussian noises (Kim et al, 1999;Ali et al, 2002;Sheikhzadeh and Deng, 1998;Gajic and Paliwal, 2006). Additionally, we find that the performance benefit is lost in babble.…”

“…The reason for this might be that their robustness varies in different types of background noise. Many studies have evaluated timing representations using a white noise interferer only (e.g., Ali et al, 2002;Sheikhzadeh and Deng, 1998). The robustness of timing features to more ecologically relevant noise backgrounds, such as multi-talker babble, is less clear, with the few studies that employ babble reporting relatively poor results (e.g., Gajic and Paliwal, 2006).…”

A frequency-selective feedback model of auditory efferent suppression and its implications for the recognition of speech in noise

“…For example, improvements could be gained by encoding the speech signal with spike timing information, instead of (or in addition to) firing rate. A number of studies (e.g., Kim et al, 1999;Sheikhzadeh and Deng, 1998;Brown et al, 2011) have shown that speech is coded more robustly by timing information in noisy conditions than by average firing rate, or other spectrally based features such as Mel frequency cepstral coefficients. If these principles are investigated further it might be possible to move the speech in noise performance of the model closer to human levels and to develop deeper insights into the underlying mechanisms.…”

Automatic speech recognition with an adaptation model motivated by auditory processing

“…A number of auditory-based approaches for noise-robust ASR have been proposed by different groups. They include models of particular psychoacoustic or physiological effects (e.g., temporal masking [11]), more comprehensive models of "effective" sound processing [15], [30], and complex, physiologically-based inner ear models [31], [14]. We chose to model a single effect, synaptic adaptation.…”