Separable spectro-temporal Gabor filter bank features: Reducing the complexity of robust features for automatic...

Area of Science:

• Speech processing
• Auditory perception
• Machine learning for ASR

Background:

• Robust feature extraction is crucial for automatic speech recognition (ASR) in noisy environments.
• Traditional spectro-temporal Gabor filter banks (GBFB) assume simultaneous processing of spectral and temporal information.
• Investigating the necessity of this simultaneous processing for ASR performance is key.

Purpose of the Study:

• To determine if spectral and temporal processing can be separated for robust ASR feature extraction.
• To introduce and evaluate separate Gabor filter bank (SGBFB) features.
• To compare the performance of SGBFB features against traditional GBFB and Mel-frequency cepstral coefficients (MFCCs).

Main Methods:

• Decomposing a spectro-temporal Gabor filter bank (GBFB) into independent spectral and temporal 1D-Gabor filter banks.
• Extracting separate Gabor filter bank (SGBFB) features using these independent banks.
• Evaluating SGBFB features on the CHiME keywords-in-noise recognition task.

Main Results:

• Spectral and temporal processing can be performed independently for robust ASR.
• SGBFB features allowed a 1.2 dB lower signal-to-noise ratio (SNR) compared to GBFB, improving word error rate by 12.8%.
• The real-time factor of spectro-temporal processing was reduced by over an order of magnitude.

Conclusions:

• Simultaneous spectro-temporal processing is not required for robust ASR feature extraction.
• SGBFB features offer significant improvements in ASR performance and computational efficiency.
• SGBFB features require a lower SNR compared to GBFB and MFCCs for human-listener-equivalent recognition performance.