Sparse representation of speech using an atomic speech modela).

Area of Science:

• Auditory Neuroscience
• Speech Processing
• Psychoacoustics

Background:

• Speech perception research commonly uses degradation algorithms like channel vocoding.
• Existing methods do not fully capture sparse time-frequency representations of speech.

Purpose of the Study:

• Introduce and validate an "atomic speech model" for generating sparse speech signals.
• Investigate the intelligibility of atomic speech under various listening conditions.
• Explore the impact of spectral maxima, binaural integration, and echo on speech reception thresholds.

Main Methods:

• Speech signals were processed using bandpass filters and undersampling.
• Each sample was reproduced as a Gaussian-enveloped tone (Gabor atom).
• Adaptive speech reception thresholds (SRTs) were measured as a function of atom rate in normal-hearing listeners.

Main Results:

• Atomic speech remained intelligible at low rates (80 atoms/sec) with 4 spectral maxima.
• Binaural presentation of non-overlapping atoms worsened SRTs compared to monaural presentation.
• A 100 ms echo delay significantly increased mean SRTs compared to shorter delays.

Conclusions:

• The atomic speech model is a valuable tool for studying speech perception mechanisms.
• Findings highlight the challenges in binaural integration and the disruptive effect of echo.
• The model's sparse representations offer new avenues for understanding auditory processing.