Audio Snippet Studio

Turns a caregiver interview recording into five candidate voice snippets and files them into the Audio Snippets sheet. Deepgram transcribes and separates the speakers; the language model only chooses which sentences to use, so every timestamp comes from the transcript rather than a guess.

1 · Deepgram API key

Stored only in this browser (localStorage), sent only to api.deepgram.com. Get one at console.deepgram.com. Deepgram supplies word-level timestamps and speaker labels — this is what makes the clip boundaries exact.

2 · Gemini API key

Used only to choose which sentences make good snippets — it receives the transcript text, never the audio. The same key you use for Headshot Studio works here. Get one at aistudio.google.com/apikey.

3 · Publishing endpoint

The Apps Script web app bound to the Caregivers-images spreadsheet. It appends the row and uploads the clips as the sheet's owner, so nobody using this page signs in to Google at all.
The shared secret the endpoint checks before it writes anything.
Not tested.
Both values are remembered in this browser only and are sent nowhere except the endpoint above.

4 · Settings

Defaults point at the existing Caregivers-images spreadsheet and its Audio Snippets tab. The Drive folder for clips is chosen by the endpoint — Test connection reports which one.
Edits here are saved in this browser and override the built-in prompt on this device only. Reset restores the built-in version. Save updated file downloads a copy of this page with your edits embedded (used when redeploying the site).

5 · New caregiver

Drop the interview audio here, or tap to choose a file
MP3 / M4A / WAV / OGG / FLAC · one interview at a time · up to ~50 MB

6 · Which voice is the caregiver?

Nothing transcribed yet.
Deepgram separated these voices. Gemini is then asked which one is the caregiver: talk time alone is unreliable, since an interviewer may well talk more than the applicant. The reason for the pick is shown below — change it if it looks wrong. Snippets are only ever taken from the selected voice.

7 · Proposed snippets

Nothing analyzed yet.
Publishing uploads one MP3 per snippet to the Drive folder, then appends a single row for this caregiver. The snippet marked Selected becomes the row's starting choice — it can be changed later in the sheet.

8 · Operations log