Dataset of directional room impulse responses for realistic speech data

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Acquiring real-world multi-channel speech data is costly and labor-intensive.
• Existing methods often use simulated Room Impulse Responses (RIRs) from simple 'shoebox' rooms, limiting realism for moving speakers.
• Far-field speech processing for applications like smart assistants requires handling moving speakers in reverberant conditions.

Purpose of the Study:

• To create a dataset of directional RIRs recorded in real environments (classroom, corridor).
• To facilitate the generation of realistic multi-channel speech data for moving speakers.
• To support research in far-field speech processing for home automation and smart assistants.

Main Methods:

• Recorded directional RIRs at multiple locations on a fine grid within a classroom and a large corridor.
• Developed a method to simulate moving speakers by generating random trajectories on the RIR grid.
• Utilized the overlap-add method to convolve monaural speech with RIRs for spatialized audio generation.

Main Results:

• A comprehensive dataset of directional RIRs from real-world acoustic spaces.
• A validated methodology for simulating moving speakers in reverberant environments.
• Demonstrated the utility of the dataset and methods with an example application.

Conclusions:

• The provided dataset significantly enhances the realism of simulated multi-channel speech data.
• Enables more robust development and testing of far-field speech processing systems for dynamic environments.
• Addresses a critical need for realistic data in speech technology research and development.