use case · multi-model
A transcript tells you what was said. Pitch tells you how — the intonation, the questioning rise, the emphasis. Here two on-device models run on the same speech: Whisper writes the words, and CREPE traces the fundamental frequency, so you can see the prosody underneath the sentence. Both local, nothing uploaded.
CREPE (~2 MB, onnxruntime-web) tracks pitch; Whisper-tiny.en (transformers.js) transcribes. Both auto-initialise from cache on a return visit.
Use the public-domain JFK inaugural sample, upload audio you have permission to process, or record your own speech:
Speech pitch contour (the prosody underneath the words):
ASR throws away pitch — Whisper gives you the words but a flat transcript can't tell a question from a statement, or where the stress fell. CREPE recovers exactly that signal. Run together they give a fuller picture of speech: transcript plus intonation — useful for language learning, expressive TTS reference, sentiment/emphasis cues, or clinical voice analysis. Each runs off the main thread in its own worker; CREPE is tiny and instant, Whisper is the heavier of the two.