Adaptation to New Microphones Using Artificial Neural Networks With Trainable Activation Functions

Area of Science:

• Speech Processing
• Machine Learning
• Artificial Intelligence

Background:

• Automatic Speech Recognition (ASR) systems require adaptation to handle variations in speech spectrum caused by microphone mismatch.
• Connectionist (hybrid) ASR systems benefit from techniques that adjust parameters using limited enrollment data.

Purpose of the Study:

• To investigate four distinct adaptation schemes for connectionist ASR systems.
• To enable ASR models to learn microphone-specific hidden unit contributions efficiently.
• To develop adaptation methods suitable for large-scale online deployment with minimal storage requirements.

Main Methods:

• Utilizing Hermite activation functions.
• Introducing bias and slope parameters in sigmoid activation functions.
• Injecting unit-specific amplitude parameters for sigmoid units.
• Combining bias/slope and amplitude parameter adjustments.

Main Results:

• Experimental results demonstrate reduced word error rates on the Wall Street Journal corpus (Spoke 6 task) compared to unadapted systems.
• Proposed adaptation schemes outperform simple multicondition training.
• Approaches show favorable comparison against linear regression-based methods, using significantly fewer parameters (up to 15 orders of magnitude less).
• Adaptation strategies remain effective even with a single adaptation sentence.

Conclusions:

• The investigated adaptation schemes offer a simple yet effective solution for improving ASR performance under microphone mismatch.
• These methods are highly desirable for deep neural networks due to their small storage footprint, facilitating online deployment.
• The proposed strategies provide a robust and parameter-efficient approach to ASR model adaptation.