Gemini API Integration

Gemini powers Ayumi’s two optional cloud transcription engines: Gemini, which reshapes a recording with your prompt, and Gemini Transcribe, Google’s dedicated speech model with language, speaker and vocabulary options. Both use the same key. You bring your own API key (BYOK), so you have full control over costs and usage.

Tip

You do not need this to start. Ayumi transcribes on-device out of the box, with no key and no account — see Voice Recording. Set up Gemini when you want a recording reshaped into a summary, a meeting note, or a written-up journal entry, rather than transcribed word for word — or when you want what the audio says beyond the words (see below).

Gemini hears more than the words

Most AI voice-note tools work in two steps: a speech-to-text engine turns the audio into words, then a language model rewrites those words. Anything that wasn’t a word — footsteps, rain, a train announcement, a child singing in the next room, the tone of your voice — is gone before the AI ever sees it.

The Gemini engine skips that first step. Ayumi sends the recording itself to Gemini together with your prompt, and Gemini, a multimodal model, listens to the audio directly. Your prompt can ask for things only the sound carries:

The audio carries	Gemini can note, for example
Background sound	Walking outside, driving with navigation prompts, rain, a busy station
Other voices	Someone else talking to you, a podcast or the radio playing — kept apart from your own words
Your voice	Whispering, hurried, tired, excited

For a voice journal, this is often what fades first: not what you said, but where you were and what the moment felt like. A dedicated transcription engine can’t give that back, because it only ever returns text. To get it, add a line to your profile’s prompt:

After the transcript, add a "## Setting" section with one or two sentences on the
scene as heard in the audio: where the speaker seems to be, background sounds,
anyone else who speaks, and the tone of voice.

Ayumi doesn’t send your location or weather to Gemini — it keeps those in the entry’s frontmatter itself. Whatever Gemini says about the setting comes from the sound and your prompt.

The setting is Gemini’s guess

A description of the scene is inferred from the sound. It can be wrong, and it can change if you transcribe the same recording again — a place name, for instance, may come out differently. Treat it as color rather than a record; the audio stays attached to the entry so you can check. Background voices are sent to Google along with yours.

This applies to the Gemini engine, which takes a prompt. Gemini Transcribe, like the on-device engine and the other transcription engines, returns a transcript only.

Getting a Gemini API Key

1. Visit Google AI Studio
2. Sign in with your Google account
3. Click Create API key, create a project for Ayumi, and create the key
4. Copy the key

Get a Gemini API Key, Step by Step shows each screen, from creating the key to turning on billing.

Caution

Keep your API key private. Never share it publicly.

How Google handles your data

What Google does with the audio and prompts Ayumi sends depends on one thing: whether the Google Cloud project behind your API key has billing enabled. It does not depend on where you created the key. These rules come from Google’s Gemini API Additional Terms of Service (last updated 2026-04-28), not from Ayumi.

	Unpaid (free tier)	Paid
When it applies	The key’s project has no active Cloud Billing account	The key’s project is linked to an active Cloud Billing account. This includes keys created in Google AI Studio
Prompts and responses	May be used to improve Google’s products and machine learning technologies. Human reviewers may read, annotate and process them	Not used to improve Google’s products. Logged for 55 days only to detect and prevent violations of Google’s Prohibited Use Policy
Google’s guidance	Don’t submit sensitive, confidential or personal information	—

The terms themselves say only “a limited period.” The 55 days comes from Google’s Abuse monitoring page (last updated 2026-06-09): prompts, any context you send with them, and the responses are kept for 55 days to detect and prevent policy violations. That page does not distinguish between paid and unpaid use, so the 55-day log applies to free-tier keys too. Data kept for this purpose is not used to train models other than those Google uses for policy enforcement.

If you are in the European Economic Area, Switzerland or the United Kingdom, the paid data terms apply even when you use the free quota.

Journal entries are often personal, so we recommend enabling billing on the project that holds your Ayumi key. If a profile uses its own Override API Key, the same applies to that key’s project. Light personal use typically costs little; see Google’s pricing for current rates.

Enabling billing in Google AI Studio

For screenshots of each step, see Turn on billing in the step-by-step guide.

1. In Google AI Studio, open the API keys page (or the Projects page).
2. Find the project your Ayumi key belongs to and click Set up billing in the Billing Tier column.
3. Choose an existing billing account, or add a new one with your contact details and payment method. If you have never used Google billing, you first select your country and accept the terms.
4. Depending on your account, Google asks you to prepay (from $5), lets you choose between Prepay and Postpay, or places you on Postpay. Once that is done, the project is paid.

Your existing key keeps working, so there is nothing to change in Ayumi. With Prepay, usage is deducted from your credit balance; when it reaches zero, Gemini requests fail until you add credits or turn on auto-reload. On-device transcription keeps working either way. See Google’s billing guide for details.

Configuring in Ayumi

macOS

1. Open Ayumi > Settings (or Cmd+,)
2. Go to the Transcription tab
3. Paste your Gemini API key under Gemini
4. Star a Gemini profile (such as Clean Transcript) to make it the default, or pick one per recording

iOS

1. Tap the More button (…) at the top right of the journal, then Settings > Transcription
2. Enter your Gemini API key under Gemini
3. Star a Gemini profile to make it the default, or pick one per recording

The key under Gemini is the provider’s base key: every Gemini and Gemini Transcribe profile uses it unless the profile sets an Override API Key of its own, for example to bill one profile to a different project. The engine, model, prompt and transcript options belong to the profile.

Keys are stored securely in your device’s Keychain and saved as you type; there is no Save button.

Supported Models

Model	Description
gemini-3.8-flash	Latest stable Flash; the default
gemini-3.6-flash	Previous stable Flash
gemini-3.5-flash-lite	Lighter model, faster and cheaper
Custom	Enter any Gemini model ID (preview or soon-to-retire models keep working until Google shuts them down)

Google now routes gemini-3.7-flash to gemini-3.8-flash, and limits the 2.5 models to accounts that already used them, so they are no longer listed. A profile that already uses one keeps it as a custom model. See Google’s model list for every available model ID, and its deprecation schedule for shutdown dates.

Using a newer model

When Google releases a model newer than the ones listed here, you can use it right away without waiting for an Ayumi update. In the profile editor, set Model to Custom and type the model ID exactly as it appears in Google’s model list (for example gemini-3.8-flash, without a models/ prefix). Pick a model that accepts audio input.

Gemini Transcribe

The built-in Gemini Transcribe profile uses Google’s dedicated speech model with automatic language detection and Smart style; star it to make it the default, or duplicate it to set the options above. It runs on the same API key and takes no prompt; its options appear in their own section:

Option	Values
Language	Automatic, or one of 85+ languages
Style	Smart (cleaned up) or Verbatim (word for word)
Label Speakers	Verbatim only; up to 8 speakers
Word Timestamps	Verbatim only; a [mm:ss–mm:ss] range per paragraph
Custom Vocabulary	A newline-separated Markdown file (for example in iCloud Drive), terms typed into the profile, or both

The model used is gemini-3.5-transcribe. See Voice Recording for how each option shapes the transcript.

Profiles

A profile bundles an engine with what it should do, so one name selects the whole configuration:

1. Go to Settings > Transcription
2. Tap Add Profile
3. Give it a name
4. Choose the Engine — On-device, Gemini, or Gemini Transcribe — and, for Gemini, the Model
5. Write your prompt instructions (Gemini), or set the language, style, speaker labels and extra vocabulary (Gemini Transcribe)
6. Save

Profiles appear in the recording view, and Shortcuts can pick them from a list. The starter profiles are ordinary profiles: edit, duplicate, reorder or delete them as you like.

Switching Back

Star the On-device profile to stop audio being sent to Google by default. Your Gemini key and profiles stay where they are, so a Gemini profile keeps working when you pick it and you can re-transcribe an entry with it at any time.

Data Privacy

When using either Gemini engine:

• Audio data is sent to Google’s servers for processing
• Ayumi shows a consent dialog before the first API call
• No data is stored on Ayumi’s servers — the API call goes directly from your device to Google
• Whether Google may use that data to improve its products depends on billing on your key’s project; see How Google handles your data and Google’s Gemini API terms