Throat and acoustic paired speech dataset for deep learning-based speech enhancement

Area of Science:

• Speech processing
• Acoustics
• Machine learning

Background:

• Capturing clear speech in high-noise environments is difficult.
• Throat microphones offer noise suppression but lose high-frequency speech details.
• Lack of standard datasets hinders deep learning for throat microphone enhancement.

Purpose of the Study:

• Introduce the Throat and Acoustic Paired Speech (TAPS) dataset.
• Provide a standard resource for developing and evaluating speech enhancement algorithms.
• Improve speech clarity and content restoration from throat microphone recordings.

Main Methods:

• Collected paired speech data from 60 Korean speakers using throat and acoustic microphones.
• Developed an optimal alignment method to reconcile signal differences.
• Evaluated three deep learning models on the TAPS dataset.

Main Results:

• The TAPS dataset facilitates research in speech enhancement.
• Mapping-based deep learning approaches showed superior performance.
• Speech quality and content restoration were significantly improved.

Conclusions:

• The TAPS dataset is a valuable resource for advancing throat microphone speech enhancement.
• Deep learning, particularly mapping-based methods, shows promise for improving throat microphone audio.
• Standardized datasets are crucial for progress in this field.