Reconstruction of speech from whispers

Area of Science:

• Speech processing
• Bioacoustics
• Assistive technology

Background:

• Whispered speech differs significantly from normal phonated speech in spectral characteristics and excitation.
• Existing speech synthesis methods are not optimized for reconstructing normal speech from whisper.
• Aphonia and vocal impairments limit communication for many individuals.

Purpose of the Study:

• To develop a real-time system for reconstructing normal speech from whispered inputs.
• To create a functional voice prosthesis for individuals with aphonia.
• To explore methods for improving verbal communication in environments where normal speech is unsuitable.

Main Methods:

• Utilizing the mixed excitation linear prediction (MELP) model for normal speech synthesis.
• Developing techniques to estimate MELP parameters from whispered speech, including spectral smoothing and formant modification.
• Proposing methods for synthesizing the appropriate excitation signal for normal speech reconstruction.
• Analyzing trade-offs between computational complexity, system delay, and reconstruction accuracy.

Main Results:

• Demonstrated feasibility of real-time normal speech synthesis from whispers.
• Proposed specific algorithms for adapting whispered speech features to the MELP model.
• Identified key differences in spectral and excitation characteristics between whispered and phonated speech.

Conclusions:

• The proposed method offers a viable approach for real-time speech reconstruction from whispers.
• This technology has significant potential as a voice prosthesis for aphonic individuals.
• Further research can optimize the system for enhanced accuracy and reduced computational load.