TSIP-Net

This repository provides an inference-only implementation of TSIP-Net, a no-reference speech intelligibility prediction model for speech affected by competing speech. It takes an audio file as input and returns one intelligibility score. Clean reference speech is not required.

Installation

Python 3.11 or newer is required. The checkpoint is tracked with Git LFS.

git lfs pull
python -m pip install -e .

Usage

tsipnet path/to/audio.wav

Use --device cpu or --device cuda:0 to select a device. The bundled checkpoint is loaded from checkpoints/ automatically.

Python API:

from tsipnet import TSIPNet

model = TSIPNet.from_pretrained()
score = model.predict("path/to/audio.wav")
print(score)

Input audio is resampled to 16 kHz when needed.

Citation

@article{wang_tsip-net_2026,
    title = {{TSIP}-{Net}: {No}-reference speech intelligibility prediction in the presence of competing speech},
    volume = {179},
    issn = {0167-6393},
    shorttitle = {{TSIP}-{Net}},
    url = {https://www.sciencedirect.com/science/article/pii/S0167639326000257},
    doi = {10.1016/j.specom.2026.103377},
    journal = {Speech Communication},
    author = {Wang, Haolan and Chan, Wai-Yip and Jensen, Jesper},
    month = apr,
    year = {2026},
    pages = {103377},
}