BenchEcho helps a maker compare observations from short, quiet bench tests without losing the recording behind each note. A trial name and changed variable are chosen before inference, so a misheard spoken label cannot silently attach an observation to the wrong experiment.
Local Whisper tiny.en turns the recording into an editable draft. The source player stays beside the text, the original AI proposal remains available, and a person must explicitly review and confirm the note. Editing a confirmed observation immediately removes it from the comparison until it is reviewed again. An exported evidence manifest includes source filenames, SHA256 hashes and reviewed versions; original audio remains separate.
AI is useful here as a starting draft—not a measurement engine. Our frozen synthetic evaluation exposed serious noise-related errors, including incorrect units and negation. We rejected the original noisy, hands-busy premise and narrowed the product to quiet post-trial review. BenchEcho does not infer trusted measurements, diagnose faults or assign pass/fail.
The prototype runs on localhost with Node, FFmpeg and a downloaded model, without an API key. The browser session is memory-only. A clean installation on a GitHub-hosted test runner, model-hash checks and eight automated tests have passed. The built-in synthetic workflow is verified; ordinary file-upload UI testing is incomplete. The video is a silent, scripted demonstration using synthetic audio.
Built With
- css
- ffmpeg
- html
- javascript
- node.js
- onnx
- transformers.js
- whisper
Log in or sign up for Devpost to join the conversation.