An API for how a call sounds.
Send a recording or live conversation and get back the transcript, speaker turns, and measured vocal delivery for each person: loudness, pitch, pace, and change from their earlier audio. Live and recorded analysis use the same API key and credit balance.
The API
POST /v1/analyze/audio for recordings and WS /v1/stream/realtime for live audio. Both produce the same turn list: text, speaker_id, and prosody fields (dBFS, F0, voicing/pause ratios, speaker-relative change). Diarization and transcription are in the same response, so one vendor covers all three.
The model
Measurements and speaker identity come from ProsodySSM, one model listening to the audio itself: language-agnostic, and stateful per speaker so “louder” means louder than that person on this call. It decides who is speaking from the voice; transcription runs alongside, so the product surface stays one key, one response.
Measurements you can audit
ProsodyAI does not label emotions, score sentiment, or tell you what a caller is feeling or intending. Those claims require labeled human judgment that this model was not trained on, so it does not make them. What it returns are measurements of the audio and their movement over time: quantities you can check against the waveform yourself. Deciding that a rise in level and pitch means a caller is escalating is your policy to set, on top of numbers that are auditable.