Automatic speaker diarization for natural conversation analysis in autism clinical trials

Area of Science:

• Neuroscience
• Speech and Language Pathology
• Computational Linguistics

Background:

• Social communication deficits are a core symptom of autism spectrum disorder (ASD).
• Current methods for assessing these deficits are time-consuming and subjective.
• Automated analysis of speech in ASD typically requires manual transcription, limiting efficiency.

Purpose of the Study:

• To evaluate the feasibility of using speaker diarization algorithms on natural conversations for objective ASD communication assessment.
• To determine if automated measures from home-based conversations can serve as sensitive indicators of treatment effects.

Main Methods:

• Applied automated speaker diarization to natural conversations of individuals with ASD and their partners in a home setting.
• Calculated the average speech duration per turn for participants.
• Correlated automated speech features with established clinical measures like the Vineland Adaptive Behaviour Scales (VABS).

Main Results:

• A significant correlation was found between average speech duration per turn and VABS expressive communication scores (r=0.51, p=7×10⁻⁵).
• Automated analysis of talkativeness from natural conversations is feasible.
• This automated measure demonstrates potential for objective evaluation.

Conclusions:

• Natural conversations can yield objective, automated measures of communication in ASD.
• Automated talkativeness measures show promise for sensitive and frequent assessment of communication challenges in ASD.
• This approach could enhance the evaluation of novel therapies for ASD.