Cadence runs six deterministic stages from the moment you drop an audio file to the moment a coaching tag is timestamped in the transcript. No stage is an unreviewed model completion. No coaching tag ships without a specific second of audio to point at.
Accept m4a, mp3, wav, flac up to 3 hours. Files are processed once, then deleted from the processing queue within 24 hours (retained only in your private library).
Two-speaker diarization with role inference (who is the rep, who is the prospect) driven by conversation structure, not filename guessing. Confidence scores below threshold are flagged, not hidden.
Every utterance is chunked into talk-track segments. Silences over 1.2 seconds are measured — silence-after-objection is a distinct signal from silence-during-feature-walkthrough.
A classification pass identifies specific conversational moments. Objections require a pattern match plus semantic match. Reference asks are detected on the prospect side only. Commitments are only fired when the rep uses future-tense first-person.
Each coaching tag is derived from a known failure mode and must point at a specific timestamp. 'Talk ratio too high' is not a tag — 'Rep talk 72% vs. best-rep avg 45%' attached to 15:00 is.
Every coaching tag becomes a shareable clip at /clip/<call>/<tag>. Re-running the same audio on a later pipeline produces a new tag set. Old clips keep resolving — the signed version is the artifact.