All projects

VSPY

Human speech can be broken down into many composite sound waves, each with its own frequency and amplitude. Much of the complexity of speech can be parameterized by extracting specific indices of these frequencies, such as their strength relative to each other, the amount of noise relative to harmonic energy, etc. These indices are often referred to as “low-dimensional acoustic features” and are extremely useful in the analysis of speech quality. VSPY is a Python package I made that extracts many perceptually motivated acoustic features from the raw speech signal. These features can then be used to assess voice quality in humans, naturalness in TTS, or to build classification models for things like speech biomarkers or deepfake detection. It’s still a work in progress, but try it out.

VSPY is a Python port of VoiceSauce, written in MATLAB. It also extracts other features using methods created by voice scholars such as Christian DiCanio and Will Styler.