Domain Adaptation with Augmented Data by Deep Neural Network Based Method Using Re-Recorded Speech for Automatic...

Area of Science:

• Speech Recognition
• Artificial Intelligence
• Machine Learning

Background:

• Automatic Speech Recognition (ASR) models, primarily based on Artificial Neural Networks (ANNs), require extensive, condition-matched training data.
• Real-world speech data acquisition across diverse acoustic environments (e.g., mobile telephony, classroom recordings) is costly and challenging.
• Existing ASR systems struggle with performance degradation due to variations in recording channels and environments.

Purpose of the Study:

• To investigate and propose effective methods for domain adaptation in ASR using simulated and augmented data.
• To reduce the cost and effort associated with acquiring diverse, real-world speech data for ASR training.
• To improve the character error rate (CERR) of ASR systems in challenging acoustic conditions, specifically targeting telephone speech via Long Term Evolution (LTE) channels.

Main Methods:

• Training ASR models with simulated augmented data, followed by fine-tuning for domain adaptation using deep neural network (DNN)-based simulated data and re-recorded data.
• Employing DNN-based feature transformation to generate realistic speech features from clean condition recordings.
• Conducting a comparative investigation of different recording channel adaptation techniques for real-world speech recognition.

Main Results:

• The proposed method achieved a 27.0% character error rate reduction (CERR) for DNN-hidden Markov model (DNN-HMM) hybrid ASR approaches.
• A significant 36.4% CERR was observed for end-to-end ASR approaches targeting the LTE channel of telephone speech.
• Simulated data augmentation and DNN-based feature transformation proved effective for domain adaptation in ASR.

Conclusions:

• Training ASR models with simulated data and fine-tuning with DNN-based techniques offers a cost-effective solution for improving performance in diverse acoustic environments.
• The proposed approach demonstrates substantial improvements in ASR accuracy for real-world telephone speech transmitted over LTE channels.
• This research provides a viable strategy for enhancing ASR robustness against acoustic variations without extensive real-world data collection.