Word timestamps
Timed words support within-recording pace, pause, duration, and spacing observations.
Loading…
Analysis methodology
SpeakPrint is an experimental replay and reflection tool. It shows what its inputs support, makes uncertainty visible, and does not infer why a speech moment occurred.
Inputs
Evidence derived from the same source does not count twice. A marker is called corroborated only when at least two independent sources contribute.
Timed words support within-recording pace, pause, duration, and spacing observations.
Provider certainty describes recognition confidence. It is not a direct measurement of speech clarity.
Literal transcript content supports conservative repetition and filler-word observations.
When decoding succeeds, local audio energy can support whether neighboring word boundaries appear separated.
Synthetic tests protect deterministic behavior; they do not prove real-world accuracy. A publishable accuracy claim requires consented human recordings, written annotation definitions, at least two independent reviewers, a held-out evaluation set, and subgroup reporting.
SpeakPrint must also receive qualified speech-language review of the sampling, labels, and public interpretation before describing the engine as clinically accurate. Until that work is complete, the product remains experimental and deliberately observational.
Comparisons are most meaningful when recordings use similar prompts, lengths, environments, and analysis versions. A change in a descriptive index is not proof of improvement or decline.
Version 1.2.0 is currently calibrated for English transcripts. A detected non-English recording is labeled limited; its timing remains visible, but language-dependent observations should not be interpreted.
Every Voiceprint stays linked to its transcript evidence, audio, and stated limits.
Try SpeakPrint