ASR-based speech intelligibility prediction: A review

Area of Science:

• Speech processing
• Machine learning
• Acoustic analysis

Background:

• Traditional speech intelligibility prediction (SIP) methods face challenges with prior knowledge requirements and limited generalization across diverse conditions.
• Automatic speech recognition (ASR) has emerged as a viable machine learning-based alternative for objective and broadly applicable SIP.

Purpose of the Study:

• To provide a comprehensive overview of ASR-based speech intelligibility prediction (SIP) research.
• To analyze and compare different ASR-based SIP methodologies, highlighting their strengths and weaknesses.
• To discuss the current state and future prospects of ASR in speech intelligibility assessment.

Main Methods:

• Reviewing and synthesizing existing literature on ASR-based speech intelligibility prediction (SIP).
• Analyzing the architectural differences and underlying principles of various ASR-based SIP models.
• Evaluating the performance and generalization capabilities of ASR approaches across different noise types, degradation levels, and speech materials.

Main Results:

• ASR-based SIP methods demonstrate significant potential in overcoming the limitations of conventional techniques.
• These approaches show promising performance across various challenging acoustic environments and speech content.
• The rapid development and deployment of ASR in SIP contexts are evident.

Conclusions:

• ASR-based methods represent a significant advancement in objective speech intelligibility prediction.
• Further research is needed to enhance generalization and address specific limitations.
• Future work should explore novel ASR architectures and their integration into advanced SIP systems.