A TOP-DOWN AUDITORY ATTENTION MODEL FOR LEARNING TASK DEPENDENT INFLUENCES ON PROMINENCE DETECTION IN SPEECH

Area of Science:

• Auditory Neuroscience
• Machine Learning
• Signal Processing

Background:

• Top-down attention models guide focus to relevant information in complex sensory environments.
• Existing models often lack biological plausibility or struggle with auditory scene analysis.

Purpose of the Study:

• To present a novel, biologically plausible top-down auditory attention model.
• To investigate task-dependent influences on auditory attention.
• To evaluate the model's performance on a prominent syllable detection task.

Main Methods:

• Extracted multi-scale features mimicking central auditory system processing.
• Converted features to low-level auditory "gist" features representing scene information.
• Trained a top-down model to map auditory gist features to scene categories.
• Tested the model on syllable prominence detection in broadcast news speech.

Main Results:

• Achieved 85.8% accuracy in prominent syllable detection at the syllable level.
• Demonstrated the model's effectiveness on the BU Radio News Corpus.
• Performance closely matched reported human capabilities for this task.

Conclusions:

• The proposed auditory attention model effectively utilizes gist features for task-dependent attention.
• The model shows strong potential for applications in speech processing and auditory scene analysis.
• Biologically plausible modeling can yield high performance in complex auditory tasks.