Using neural network and principal component analysis to study vowel recognition with temporal envelope cues.

Area of Science:

• Speech processing
• Auditory perception
• Machine learning in acoustics

Background:

• Speech recognition heavily relies on temporal envelope cues in both normal-hearing and cochlear-implant users.
• Understanding the specific features within temporal envelopes is key to improving auditory prosthetics and speech recognition algorithms.

Purpose of the Study:

• To investigate the role of temporal envelope features in vowel recognition using principal component analysis (PCA).
• To develop and evaluate a feedforward artificial neural network model for vowel recognition based on extracted temporal envelope cues.

Main Methods:

• Extracted temporal envelopes from 1 to 8 spectral bands for twelve vowels spoken by 30 speakers.
• Applied PCA to identify key components within the temporal envelopes.
• Constructed a 3-layer feedforward artificial neural network to assess vowel recognition accuracy.

Main Results:

• Achieved 63% correct vowel recognition using only 4-band temporal envelope cues.
• Identified phonemic transition cues and steady-state amplitude cues as critical components for high recognition accuracy.
• PCA effectively reduced feature dimensionality while retaining essential information for vowel identification.

Conclusions:

• Temporal envelope cues, particularly phonemic transitions and steady-state amplitudes, are highly informative for vowel recognition.
• The findings support the development of advanced algorithms for automatic speech recognition and improved auditory prosthetic devices.