TTFNet: Temporal-Frequency Features Fusion Network for Speech Based Automatic Depression Recognition and Assessment.

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Depression diagnosis lacks objective, convenient early screening methods.
• Phonological differences exist between depressed patients and healthy individuals.

Purpose of the Study:

• To propose an automatic depression detection method using hybrid speech features.
• To enhance early depression screening through advanced deep learning techniques.

Main Methods:

• Developed TTFNet, a deep learning model utilizing hybrid speech features.
• Employed quaternion representation for frequency domain features (log-Mel spectrogram).
• Integrated quaternion VisionLSTM and sLSTM with wav2vec 2.0 for temporal features.
• Utilized an XConformer block and dual-path fusion for feature complementarity.

Main Results:

• TTFNet outperforms current state-of-the-art methods on multiple datasets (AVEC 2013, AVEC 2014, DAIC-WOZ, E-DAIC).
• Achieved superior performance in both depression recognition and severity prediction tasks.

Conclusions:

• The proposed TTFNet method offers a promising approach for objective and convenient depression screening.
• Hybrid speech features and advanced deep learning architectures significantly improve depression detection accuracy.