Machine learning based sample extraction for automatic speech recognition using dialectal Assamese speech

Area of Science:

• Speech Processing and Recognition
• Machine Learning Applications
• Human-Computer Interaction

Background:

• Automatic Speaker Recognition (ASR) is evolving with Human-Computer Interaction (HCI), incorporating big data and the Internet of Things (IoT).
• Learning-based techniques are gaining traction in ASR due to their ability to mimic biological behavior for improved modeling and processing.
• Current ASR methods are integrating big data and IoT concepts, necessitating advanced machine learning approaches.

Purpose of the Study:

• To investigate machine learning (ML) approaches for extracting relevant samples from big data for ASR.
• To apply soft computing techniques for ASR of Assamese speech, addressing dialectal variations.
• To evaluate the performance of various ML models, including Artificial Neural Networks (ANNs) and Deep Neural Networks (DNNs), for speaker recognition tasks.

Main Methods:

• Utilized ML techniques, including feedforward (FF) and deep neural networks (DNNs), for sample extraction and feature engineering.
• Employed Multi-Layer Perceptrons (MLPs) with raw speech, extracted features, and frequency domain inputs for class information learning.
• Applied Recurrent Neural Networks (RNNs) and Fully Focused Time Delay Neural Networks (FFTDNNs) with spectral and prosodic features for recognition.

Main Results:

• Proposed ML-based sentence extraction techniques and a composite feature set with RNN as a classifier outperformed other methods.
• ANNs in feedforward form were used as feature extractors, enabling performance evaluation and comparison.
• Experimental results demonstrated that leveraging big data samples significantly enhanced ASR system learning.

Conclusions:

• ANN-based sample and feature extraction techniques are efficient for integrating ML into big data aspects of ASR systems.
• The developed ML approaches are effective for recognizing speaker, dialect, gender, and mood variations in Assamese speech.
• The study highlights the potential of big data and advanced ML techniques for advancing ASR capabilities.