Modification of computational auditory scene analysis (CASA) for noise-robust acoustic feature

Minseok Kwon, Purdue University

Abstract

While there have been many attempts to mitigate interferences of background noise, the performance of automatic speech recognition (ASR) still can be deteriorated by various factors with ease. However, normal hearing listeners can accurately perceive sounds of their interests, which is believed to be a result of Auditory Scene Analysis (ASA). As a first attempt, the simulation of the human auditory processing, called computational auditory scene analysis (CASA), was fulfilled through physiological and psychological investigations of ASA. CASA comprised of Zilany-Bruce auditory model, followed by tracking fundamental frequency for voice segmentation and detecting pairs of onset/offset at each characteristic frequency (CF) for unvoiced segmentation. The resulting Time-Frequency (T-F) representation of acoustic stimulation was converted into acoustic feature, gammachirp-tone frequency cepstral coefficients (GFCC). 11 keywords with various environmental conditions are used and the robustness of GFCC was evaluated by spectral distance (SD) and dynamic time warping distance (DTW). In "clean" and "noisy" conditions, the application of CASA generally improved noise robustness of the acoustic feature compared to a conventional method with or without noise suppression using MMSE estimator. The intial study, however, not only showed the noise-type dependency at low SNR, but also called the evaluation methods in question. Some modifications were made to capture better spectral continuity from an acoustic feature matrix, to obtain faster processing speed, and to describe the human auditory system more precisely. The proposed framework includes: 1) multi-scale integration to capture more accurate continuity in feature extraction, 2) contrast enhancement (CE) of each CF by competition with neighboring frequency bands, and 3) auditory model modifications. The model modifications contain the introduction of higher Q factor, middle ear filter more analogous to human auditory system, the regulation of time constant update for filters in signal/control path as well as level-independent frequency glides with fixed frequency modulation. First, we scrutinized performance development in keyword recognition using the proposed methods in quiet and noise-corrupted environments. The results argue that multi-scale integration should be used along with CE in order to avoid ambiguous continuity in unvoiced segments. Moreover, the inclusion of the all modifications was observed to guarantee the noise-type-independent robustness particularly with severe interference. Moreover, the CASA with the auditory model was implemented into a single/dual-channel ASR using reference TIMIT corpus so as to get more general result. Hidden Markov model (HTK) toolkit was used for phone recognition in various environmental conditions. In a single-channel ASR, the results argue that unmasked acoustic features (unmasked GFCC) should combine with target estimates from the mask to compensate for missing information. From the observation of a dual-channel ASR, the combined GFCC guarantees the highest performance regardless of interferences within speech. Moreover, consistent improvement of noise robustness by GFCC (unmasked or combined) shows the validity of our proposed CASA implementation in dual microphone system. In conclusion, the proposed framework proves the robustness of the acoustic features in various background interferences via both direct distance evaluation and statistical assessment. In addition, the introduction of dual microphone system using the framework in this study shows the potential of the effective implementation of the auditory model-based CASA in ASR.

Degree

Ph.D.

Advisors

Talavage, Purdue University.

Subject Area

Electrical engineering|Acoustics

COinS