HypernasalityNet: Deep recurrent neural network for automatic hypernasality detection.

Area of Science:

• Speech-language pathology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Cleft palate patients often exhibit hypernasal speech due to inadequate velopharyngeal closure.
• Current hypernasal speech assessment relies on subjective evaluations by speech-language pathologists.
• Automated detection systems can aid clinicians in diagnosing and treating hypernasal speech.

Purpose of the Study:

• To develop a Long Short-Term Memory (LSTM) based Deep Recurrent Neural Network (DRNN) system for automatic hypernasal speech detection.
• To explore the feature mining and classification capabilities of the LSTM-DRNN system.
• To provide an aided diagnostic tool for cleft palate patients with hypernasality.

Main Methods:

• Utilized 14,544 Mandarin vowel recordings from 144 children (72 with hypernasality, 72 controls, aged 5-12).
• Proposed an LSTM-DRNN system to leverage short-time dependencies in speech.
• Analyzed vocal tract features and hypernasality-sensitive vowels (/a/, /i/, /u/) for the first time.

Main Results:

• The LSTM-DRNN system achieved a highest detection accuracy of 93.35%.
• The system demonstrated superior performance compared to shallow classifiers by mining features across time and network depth.
• Vowels /i/ and /u/ were identified as the most sensitive to hypernasal speech.

Conclusions:

• LSTM-DRNN exhibits robust feature mining and classification abilities for hypernasal speech detection.
• This marks the first application of LSTM-DRNN for automatic hypernasality detection in cleft palate speech.
• Deep learning holds significant potential for aiding pathologists in speech detection.