Automating Speech Audiometry in Quiet and in Noise Using a Deep Neural Network

Area of Science:

• Audiology
• Speech Science
• Artificial Intelligence in Healthcare

Background:

• Comprehensive hearing evaluations include speech understanding assessment.
• Current speech audiometry requires time-consuming manual scoring by professionals.
• Automated scoring methods are needed to improve efficiency.

Purpose of the Study:

• To develop and validate an automated speech recognition (ASR) system for phonetic-level scoring of speech audiometry.
• To assess the performance and reliability of the ASR system in clinical settings.
• To compare the ASR system's scoring accuracy and test-retest reliability against manual scoring.

Main Methods:

• Developed a deep neural network-based automated speech recognition (ASR) system.
• Trained the ASR system using French speech materials (Lafon's cochlear lists, Dodelé logatoms).
• Evaluated the ASR system's performance and reliability with normal-hearing and hearing-impaired listeners in quiet and noisy conditions.

Main Results:

• The ASR system demonstrated statistically similar performance to manual scoring by expert hearing professionals.
• Automated scoring accuracy was consistent in both quiet and noisy listening conditions.
• The test-retest reliability of the automated scoring closely matched that of manual scoring.

Conclusions:

• The developed deep neural network-based ASR system provides an accurate and reliable method for speech audiometry.
• This automated system is validated for use in both clinical practice and research for hearing evaluations.
• The ASR system offers an efficient alternative to manual scoring, improving the assessment of speech understanding.