Suppressing aliasing noise in the speech feature domain for automatic speech recognition

Area of Science:

• Speech processing
• Signal analysis
• Machine learning

Background:

• Low-pass filtering is standard for preventing aliasing in audio signals.
• Aliasing noise can still occur in the speech feature domain, impacting recognition.
• High modulation frequencies in speech features can lead to aliasing.

Purpose of the Study:

• To identify and address aliasing noise in the speech feature domain.
• To propose a method for suppressing aliasing noise in speech features.
• To evaluate the impact of antialias processing on speech recognition performance.

Main Methods:

• Spectral analysis of speech feature streams to detect aliasing noise.
• Development and implementation of an antialias processing method for speech features.
• Experimental evaluation using large vocabulary speech recognition systems.

Main Results:

• Aliasing noise was confirmed to exist in speech feature streams.
• The proposed antialias processing method effectively suppresses this noise.
• Antialias processing improved speech recognition accuracy, especially for noisy speech data.

Conclusions:

• Antialias processing in the speech feature domain is crucial.
• This method offers significant benefits for speech recognition systems.
• The technique is particularly effective in improving the robustness of speech recognition to noise.