aminEdraki / py-intelligibility

Python implementation of a few speech intelligibility prediction algorithms

py-intelligibility

Python implementation of the following speech intelligibility prediction methods: weighted Spectro-Temporal Modulation Index (wSTMI) Spectro-Temporal Glimpsing Index (STGI)

Usage

The functions wstmi and stgi take three inputs:

d = pywstmi(clean_speech, degraded_speech, sampling_frequency)
d = pystgi(clean_speech, degraded_speech, sampling_frequency)

• clean_speech: A numpy array containing a single-channel clean (reference) speech signal.
• degraded_speech: A numpy array containing a single-channel degraded/processed speech signal.
• sampling_frequency: The sampling frequency of the input signals in Hz.

Note that the clean and degraded speech signals must be time-aligned and of the same length. The algorithms only support 10 KHz sampling rate.

Missing Features

Voice activity detection is not implemented yet.

References

If you use pywstmi or pystgi, please cite the references [1] and [2], respectively:

[1] A. Edraki, W.-Y. Chan, J. Jensen, & D. Fogerty, “Speech Intelligibility Prediction Using Spectro-Temporal Modulation Analysis,” IEEE/ACM Trans. Audio, Speech, & Language Processing, vol. 29, pp. 210-225, 2021.
[2] A. Edraki, W.-Y. Chan, J. Jensen, & D. Fogerty, “A Spectro-Temporal Glimpsing Index (STGI) for Speech Intelligibility Prediction," Proc. Interspeech, 5 pages, Aug 2021.