Langdrift
ActiveDoes the translation carry the same weight?
An open source tool that measures how differently an original and its translation land in a listener's head. It compares how the two activate the brain and returns a drift score, so you can catch translations that are technically right but feel off.
What it does
Langdrift uses Meta FAIR's TRIBE v2 model, trained on fMRI data from 720 people, to map audio onto roughly 20,000 points across the cortex. It then scores the difference between original and translated activation with a Pearson correlation. Run langdrift run original.mp3 translation.mp3 and get a number back: the higher the drift, the bigger the cognitive gap.
Key features
- CLI comparison of original versus translated audio.
- Threshold presets: hard (10% / 5%), medium (20% / 10%), low (30% / 20%).
- Supports .mp3, .wav, .m4a, .flac, .ogg, .aac and .opus.
- JSON output for automation and CI pipelines.
- One-time setup wizard for Python, Modal and HuggingFace credentials.
Stack & status
A TypeScript / Node.js CLI over a Python backend running on Modal (A10G GPU). Core model is Meta FAIR TRIBE v2 (Wav2Vec-BERT 2.0 encoder), with Pearson correlation for scoring. Apache 2.0 licensed (model weights CC BY-NC 4.0), in active development.