Changelog

What shipped in each release.

AudioLens doesn't carry a version number today — see Licenses and the rest of these docs for "Open source" as the only label used. This page tracks what's shipped and what's in development instead of version numbers.

AudioLens

AudioLens extends the original ECHO project with a substantial set of new capabilities, currently in development:

Asynchronous job execution (submit → poll → fetch result, instead of synchronous inference)
A job lifecycle API (/jobs, /jobs/{id}, /jobs/{id}/result)
An object storage abstraction (local filesystem or S3-compatible backend)
Dataset EDA (exploratory data analysis) tooling
Embedding analytics — 2D/3D projection, clustering, nearest-neighbour retrieval
Saliency faithfulness checking (deletion-curve validation of saliency maps)
Linguistic-vs-acoustic influence sweeps
Layer-wise probing (linear probes per model layer)
Jacobian Lens — a decoder-only, position-resolved lens over Whisper's decoder layers
Accent and fairness slicing across speaker/accent/language/metadata groups
Custom model ingestion (any compatible Hugging Face speech model)
Two new built-in datasets: LibriSpeech-1000 and SAVEE (SAVEE requires the user's own licensed copy — see Licenses)
Dataset fetch scripts for RAVDESS, the Speech Accent Archive, L2-ARCTIC and Common Voice, plus a wrapper for the no-account ones — datasets now come from their official sources into Backend/data instead of being described as bundled (see CLI & scripts)

[1.0.0] — 2024-10-13

LIT for Voice — initial release.

React frontend, FastAPI backend
Whisper and Wav2Vec2 model support
Redis-backed caching
Common Voice and RAVDESS sample datasets

This release was later renamed ECHO, and is now developed further as AudioLens. See Contributing for the full lineage and credits.

Was this page helpful?

AudioLens docs · MIT LicenseWaveform to logits