FAQ

FAQ

Does Arachne use an LLM to score? No. Scoring is deterministic signal processing and explicit rules. The same input and config give identical output.

What does a z-score mean here? A robust (median-based) standardisation of a deviation within one recording pair, with a floor. It is not a universally calibrated judgment. Every rule also needs a minimum absolute change.

Can it tell me my delivery is "bad"? It tells you where and how your delivery differs from a chosen reference, by rule. It cannot say that the difference is wrong, intended, or why it happened.

Why does it need the transcript and a reference? Matched text makes word-by-word comparison possible. The transcript must match the reference's words.

Why "insufficient evidence"? Because withholding a score is better than inventing one when audio is too noisy, too little speech is voiced, or alignment fails.

Different speakers? Pitch is expressed in semitones relative to each speaker's own median and scaled by their spread; energy is median-centred. Raw cross-speaker values are never compared. A different-speaker near-perfect mirror is part of the test set.

Why are the dataset and results synthetic? To validate the pipeline with exact ground truth while real recordings were not yet available. Real data is the top remaining task.

Is WhisperX supported? The backend is implemented but untested. The default aligner needs no model download.

Does it work with noisy recordings? Not reliably at about 20 dB SNR or lower. See Stress Testing.

Where's the UI? The dashboard is in progress and is built against the API. See API Reference.

How do I cite or credit AI use? The README and Devpost Built With disclose that Claude (Anthropic) assisted with code, tests and docs.