Identification of prosodic attitudes by a temporal recurrent network

Area of Science:

• Computational neuroscience
• Linguistics
• Speech processing

Background:

• Human speech utilizes fundamental frequency (F0) modulation to convey prosodic attitudes.
• Decoding these prosodic attitudes from speech is crucial for natural language understanding.
• Previous research has explored various acoustic cues, but the temporal dynamics of F0 remain a key area of investigation.

Purpose of the Study:

• To investigate the role of temporal F0 structure in discriminating prosodic attitudes.
• To develop and evaluate a temporal recurrent neural network (TRN) for prosodic attitude classification.
• To compare the TRN's performance against human subject performance in a similar discrimination task.

Main Methods:

• A temporal recurrent neural network (TRN), inspired by primate frontostriatal neurophysiology, was employed.
• The TRN processed population coding of continuous, time-varying fundamental frequency (F0) values from natural language sentences.
• The model was trained and tested on an experiment involving discrimination of six prosodic attitudes.

Main Results:

• The TRN model achieved 82.52% accuracy in discriminating six prosodic attitudes.
• Human subjects achieved 72.8% accuracy in the same discrimination task.
• The results indicate that F0 temporal structure contains significant information for prosodic attitude classification.

Conclusions:

• Fundamental frequency (F0) variations over time are highly informative for identifying prosodic attitudes in speech.
• The developed temporal recurrent neural network (TRN) demonstrates effective categorical sensitivity to these F0 dynamics.
• The TRN model shows potential as a computational tool for understanding prosodic perception and classification.