Loading page
What AudioLens is and why it exists.
AudioLens is an interpretability and debugging workbench for speech models — the audio-domain counterpart to Google PAIR's LIT (Learning Interpretability Tool), which targets text and tabular models. It lets a researcher or developer follow a single prediction from raw waveform to emitted token, using saliency maps, attention visualization, embedding projections, perturbation testing, layer probing, and fairness slicing — all on one shared timeline.
Speech models are hard to debug from the outside. A transcript or a confidence score tells you what a model predicted, not why — which parts of the waveform it attended to, whether a saliency map can be trusted, whether performance holds up across accents or speakers, or where in the network a property like pitch or phoneme identity is actually represented. AudioLens exists to make those questions answerable interactively, on real audio, without writing one-off analysis scripts for every question.
Researchers and engineers working on speech recognition or speech emotion models who need to inspect a specific prediction, compare checkpoints, stress-test robustness with perturbations, or check whether a model's errors are concentrated in particular speaker groups.
AudioLens is built on and extends ECHO, an MIT-licensed audio interpretability tool originally released as LIT for Voice (v1.0.0, 2024-10-13) and created by Anas Hussaindeen, Chandupa Ambepitiya, and Dewmike Amarasinghe. AudioLens is inspired throughout by Google PAIR's LIT. See /docs/licenses for the full license and attribution.