A novel objective function for improved phoneme recognition using time-delay neural networks

Area of Science:

• Speech Recognition
• Artificial Intelligence
• Machine Learning

Background:

• Traditional speech recognition models often use mean-squared-error (MSE) or cross-entropy (CE) objective functions for training.
• These functions aim to minimize discrepancies between predicted and actual outputs, which may not be optimal for classification tasks.

Purpose of the Study:

• To introduce and evaluate a novel objective function, the classification figure of merit (CFM), for training time-delay neural networks (TDNNs).
• To assess the effectiveness of CFM in improving the accuracy of single-speaker and multispeaker recognition of voice-stop consonants (/b,d,g/).

Main Methods:

• Implemented time-delay neural networks (TDNNs) for speech recognition tasks.
• Developed and applied a new objective function, the classification figure of merit (CFM), during TDNN training.
• Compared CFM against traditional MSE and cross-entropy (CE) objective functions, incorporating a simple arbitration mechanism.

Main Results:

• The CFM objective function demonstrated a median 30% reduction in misclassifications compared to TDNNs trained solely with the MSE objective function.
• CFM focuses on maximizing the difference for incorrect classifications, unlike MSE and CE which minimize overall error.

Conclusions:

• The classification figure of merit (CFM) offers a more effective approach for training TDNNs in speech recognition compared to traditional methods.
• This enhancement leads to significant improvements in the accurate recognition of challenging speech sounds like voice-stop consonants.