gabrielmittag/NISQA

Built on deep learning with modular architecture (CNN/feedforward → self-attention/LSTM → pooling layers), NISQA provides both single-ended quality assessment for transmitted speech and TTS naturalness prediction, with v2.0 adding multidimensional breakdowns (noisiness, coloration, discontinuity, loudness). Supports end-to-end training, finetuning, and transfer learning via PyTorch, configured through YAML files, plus includes a 14,000+ sample corpus with subjective quality labels for model development.

917 stars. No commits in the last 6 months. Available on PyPI.