Active listening.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Speech Processing

Background:

• Active inference unifies perception and action to maximize model evidence.
• Speech recognition involves processing continuous acoustic signals into discrete words.
• Active vision principles are adapted for covert speech segmentation.

Purpose of the Study:

• Introduce active listening as a unified framework for speech synthesis and recognition.
• Develop a generative model for speech, linking acoustic signals to lexical, prosodic, and speaker attributes.
• Frame speech segmentation as selecting internal actions to place word boundaries.

Main Methods:

• Described a generative model simulating acoustic signal generation from word attributes and vice versa.
• Modeled speech segmentation as selecting internal actions (word boundary placement).
• Validated the framework using simulated speech recognition and by linking to neurophysiological responses (mismatch negativity, P300).

Main Results:

• Simulated speech recognition demonstrated that inferred sentence content depends on prior beliefs and noise.
• Word boundaries were selected to maximize evidence for the internal word generation model.
• Neuronal responses correlate with belief updating, especially when prior beliefs are inaccurate.

Conclusions:

• Active listening provides a unified framework for understanding speech perception and synthesis.
• The model highlights the role of generative models and belief updating in speech processing.
• This framework has implications for understanding cognitive and neural mechanisms of speech recognition.