Intoxicated Speech Detection: A Fusion Framework with Speaker-Normalized Hierarchical Functionals and GMM...

Area of Science:

• Speech processing
• Computational paralinguistics
• Machine learning for behavioral analysis

Background:

• Speech signals contain rich paralinguistic information beyond linguistic content.
• Speaker state, including temporary physiological changes like alcohol intoxication, is difficult to detect computationally.
• Previous research highlights the challenge of accurately identifying speaker states from acoustic features.

Purpose of the Study:

• To develop and evaluate a robust system for detecting speaker state, specifically alcohol intoxication, from speech.
• To explore the effectiveness of fusing multiple feature representations (prosodic and spectral) for improved detection accuracy.
• To investigate methods for normalizing speaker variability to enhance system performance.

Main Methods:

• Utilized multiple representations of prosodic and spectral features.
• Employed classifier fusion techniques to combine individual model predictions.
• Investigated speaker normalization strategies, including using held-out baseline data.
• Evaluated system performance on the Alcohol Language Corpus at the Interspeech 2011 Intoxication Subchallenge.

Main Results:

• The fused system achieved an unweighted average recall (UAR) of 68.8% on the test set, outperforming the challenge baseline by 5.5% absolute.
• Speaker normalization significantly improved system performance, comparable to other normalization techniques.
• Matched-prompt training led to more consistent results, with the combined system achieving a UAR of 71.4%.

Conclusions:

• Fusing diverse feature representations and employing speaker normalization are crucial for accurate speaker state detection.
• The developed system demonstrates a significant advancement in identifying alcohol intoxication from speech.
• The findings offer a practical approach for building robust speaker state detection systems.