Opening up the black box
of speech models.
AudioLens is an open-source workbench for looking inside speech models, built in the open as an extension of ECHO.
Speech models are hard to debug from the outside. A transcript or a confidence score tells you what a model predicted, not why.
AudioLens makes the why answerable: which parts of the waveform the model leaned on, whether a saliency map can be trusted, whether it holds up across accents and speakers, and where in the network a property like pitch or phoneme identity actually lives. All of it interactive, on real audio, without a one-off script for every question.
It is the audio counterpart to Google PAIR's LIT, and it builds on ECHO, an MIT-licensed audio interpretability tool first released as LIT for Voice (v1.0.0, October 2024). In 2026, Januda Lelwala, Janith Mahanama and Hesandi Mallawarachchi extended it into the workbench you see here, with nine analysis panels on one shared timeline. Full attribution, including ECHO's original authors, is on the licenses page.
AudioLens is built in the open, and there are a few ways in. To propose a change, start with the contributing guide, which covers the branch and pull request conventions.
If you find a security issue, please report it privately through the security page rather than a public issue. Everything else, the code, the issues and the releases, lives on GitHub.