Voice Recording & Transcription

Ayumi lets you record voice memos and automatically transcribe them into journal entries. Transcription runs on-device by default, with Gemini available when you want the recording reshaped rather than transcribed, and dedicated cloud transcript engines from Google, Meta, xAI, and Groq when you want provider-specific speech features.

Recording Audio

From the App

1. Navigate to the recording view
2. Tap the Record button to start
3. A timer shows the recording duration
4. Tap Stop to finish recording

Microphone & Input Device

Ayumi shows which microphone is active as a badge above the recording controls, both on the ready screen and while recording. External accessories — Bluetooth, USB, wired, and car audio — are preferred automatically over the built-in mic, matching how iOS routes audio.

Bluetooth microphones such as the DJI Mic are supported. If a device pairs or disconnects, the badge updates in real time, so you can confirm the right mic before you start.

Input Level Meter

A segmented level meter fills as you speak, turning from blue to orange to red as the input approaches clipping. Use it to confirm the mic is picking you up and to keep your level out of the red. The meter also appears on the Dynamic Island and Lock Screen Live Activity.

Quick Action (iOS)

Long-press the Ayumi app icon and select Record Audio to start recording immediately.

Apple Watch

You can also record on your wrist and let your iPhone transcribe it. See Apple Watch.

Recording Status (iOS)

When recording starts — especially from a Shortcut — Ayumi shows a status screen with the active input device and elapsed time. Tap Open Recorder to jump to the full recording view, or Hide to collapse it to a compact banner that floats at the bottom of the journal list. Tap the banner any time to return to the recorder.

Dynamic Island (iOS)

While recording, a Live Activity appears in the Dynamic Island and on the Lock Screen showing the recording status, input device, input level, and duration. A distinct Starting… phase is shown while the audio session warms up, before the timer begins. You can stop recording directly from there.

Technical Details

• Audio format: AAC encoding at 44.1 kHz (mono)
• Files are stored in the entry’s assets folder

Transcription

After recording, the audio is queued for transcription. Ayumi has six engines. Which one a recording uses is decided by its profile (see below); the starred default profile is what a recording uses when nothing else chooses.

	On-device (Apple)	Gemini	Gemini Transcribe	Muse Voice Transcribe	xAI Grok Speech to Text	Groq Whisper
Setup	None	Gemini API key	Gemini API key	Meta API key	xAI API key	Groq API key
Where the audio goes	Never leaves the device	Sent to Google	Sent to Google	Sent to Meta	Sent to xAI	Sent to Groq
Network	Once, to download the language	Every transcription	Every transcription	Every transcription	Every transcription	Every transcription
Output	Verbatim transcript with a title	Whatever your prompt asks for	Transcript with a title, smart or verbatim	Transcript with a title	Transcript with a title	Transcript with a title
Prompt	Not used	Used	Not used	Not used	Not used	Not used
Custom vocabulary	No	Via the prompt	Yes, from a file or typed in	Yes, from a file or typed in	Yes, up to 100 key terms	Whisper context from a file or typed in
Speaker labels	No	Via the prompt	Yes, up to 8	Yes, 20+	Yes	No
Max input	—	~1 hour	1 hour (30 min with labels)	10 minutes	500 MB per file	25 MB per file

On-device Transcription

This is what a fresh install uses — the built-in On-device profile is the default — so your first recording becomes an entry without any setup at all. Ayumi transcribes with Apple’s speech models in your device’s language and writes a verbatim transcript, using the opening sentence as the entry title.

The speech model for your language downloads the first time you transcribe in it. Every recording after that works with no network — on a plane, underground, or in airplane mode.

Note

This engine transcribes; it does not rewrite. Profile prompts and prompt overrides do not apply, and the prompt fields are hidden while it is selected.

If your device language is not one Apple’s speech models cover, Ayumi tells you instead of transcribing it with the wrong language. Use a Gemini profile for those languages.

Gemini Transcription

Gemini sends the recording to Google and returns whatever your prompt asks for — a cleaned-up transcript, a meeting summary, a reflective entry. It needs an API key you bring yourself. See Gemini API Integration.

Ayumi asks for your consent the first time a recording is about to be sent to Google. On-device transcription never triggers that prompt.

Gemini Transcribe

Gemini Transcribe uses Google’s dedicated speech model (gemini-3.5-transcribe) with a Gemini API key. It takes no prompt — like the on-device engine it writes a transcript with the opening sentence as the title — but it gives you control over the transcript itself. The starter Gemini Transcribe profile uses automatic language detection and Smart style; edit it, or duplicate it, to set the options:

Option	What it does
Language	Automatic detects the language as it goes, including switching languages mid-recording. Pick one of the 85+ supported languages only when detection keeps guessing wrong.
Style	Smart drops fillers and repetitions and applies light formatting. Verbatim keeps every word.
Label Speakers	Verbatim only. Writes one paragraph per speaker turn (**Speaker 1:** …) for up to 8 speakers. Ayumi requests word timing alongside, which is what carries the speaker of each word; Google notes this can cost a little accuracy.
Word Timestamps	Verbatim only. Puts a [mm:ss–mm:ss] range on each paragraph — per speaker turn with Label Speakers, per sentence without. Same word-timing request and accuracy note as speaker labels.
Custom Vocabulary	Names, jargon and product names the model should recognise.

Custom vocabulary belongs to the profile and comes from two places, both sent together:

• A Markdown file — in the profile editor, tap Choose Markdown File… and pick a newline-separated list, typically kept in iCloud Drive so every device shares it. Bullets, numbering, headings and checkboxes are stripped, so a note maintained in Obsidian works as-is. Ayumi keeps secure access to the file and re-reads it for every transcription, so edits take effect immediately.
• Terms entered in the profile — one per line, for quick additions on the device you have in hand.

Note

The API accepts up to 1,000 terms but recognises best with around 100. It also cannot bias vocabulary in the same request as speaker labels or word timestamps, so while either is on the vocabulary is left out — the profile editor tells you when that is the case.

Recordings can be up to an hour long, or 30 minutes with speaker labels or word timestamps. Audio is sent to Google, so the same consent prompt as Gemini applies.

Muse Voice Transcribe

Meta’s speech model, over the Meta Model API with your own key. Like Gemini Transcribe it takes no prompt and writes a titled transcript, with Label Speakers, Turn Timestamps, language hints and keywords set on the profile — and unlike Gemini Transcribe, keywords work alongside speaker labels. Recordings are limited to 10 minutes. See Muse Voice Transcribe.

xAI Grok Speech to Text

xAI’s file transcription API, using your own xAI key. It takes no prompt and writes a titled transcript. An xAI Grok profile can automatically detect the language, format numbers and units when a language is selected, label speakers, add time ranges, preserve filler words, and bias up to 100 names or terms. Ayumi sends its M4A recording directly and accepts files up to the API’s 500 MB limit. See xAI Grok Speech to Text.

Groq Whisper

GroqCloud’s Whisper file transcription API, using your own Groq key. Choose Whisper Large V3 Turbo for the fast default or Whisper Large V3 for the accuracy-oriented option. A Groq profile can hint the language, add segment timestamps, and bias names or terms through Whisper’s context prompt. Direct uploads are limited to 25 MB. See Groq Speech to Text.

Choosing an Engine per Recording

The Profile picker on the recording screen chooses the engine and its settings. The Model picker below it overrides the engine for that recording only — it defaults to Use Profile — and each engine’s own settings can be overridden for that recording too: the prompt for Gemini, and the options under Transcript Options for Gemini Transcribe, Muse Voice Transcribe, xAI Grok Speech to Text, and Groq Whisper. The profile itself is never changed from there. The same overrides are available when you re-transcribe an entry.

You can also transcribe the same audio again later with the other engine. The recording stays attached to the entry, and each re-transcription creates a new entry rather than overwriting the original — so capturing on-device first and running it through Gemini afterwards costs you nothing.

Profiles

A profile is a named bundle of an engine and what it should do — a prompt for Gemini, or transcript options for a dedicated transcription engine. Every profile pins an engine; there is no separate default engine. The starred profile is the default, and what the Watch and background recordings use. Ayumi includes 8 built-in profiles:

Profile	Engine	Description
On-device	On-device (Apple)	Verbatim transcript, never leaves the device. The default on a fresh install
Clean Transcript	Gemini	Produces a clean, readable transcript of the audio
Meeting Summary	Gemini	Generates a structured summary with key points
Journal Reflection	Gemini	Creates a reflective journal entry from the recording
Gemini Transcribe	Gemini Transcribe	A smart transcript with automatic language detection
Muse Voice Transcribe	Muse Voice Transcribe	A transcript with automatic language detection
xAI Grok Speech to Text	xAI Grok Speech to Text	A transcript with automatic language detection
Groq Whisper (Fast)	Groq Whisper Large V3 Turbo	A fast transcript with automatic language detection

The starter Gemini profiles use gemini-3.7-flash. Every profile, starters included, can be edited, duplicated from its menu, reordered or deleted. Profiles are listed in the order you arrange them — drag one, or use Move Up / Move Down from its menu — and that order is what the recording screen and Shortcuts show.

You can also create your own profiles.

In the profile editor, choose any of the six engines, and selecting that profile — on the recording screen, in a Shortcut, or for a Watch recording — chooses the engine with it. A Gemini profile holds its model and prompt. Each dedicated cloud transcript profile holds its own supported language, timestamp, speaker, and vocabulary settings. Provider keys are set once under Gemini, Meta, xAI Grok, or Groq and shared by their profiles; a profile can override its provider key with one of its own. Consent is also provider-specific.

Note

A profile pinned to on-device is refused on a device that cannot transcribe on-device, rather than silently falling back to Gemini. Pick another profile or set a model override for that recording.

Prompt Variables

Prompts — both profiles and the per-recording Prompt Override field — can include variables that are filled in with context when the recording is submitted for transcription. Use the Insert Variable menu next to the prompt field to add them at the cursor.

Variable	Expands to
{{recent_entries}} / {{recent_entries:N}}	The full text of your latest journal entries (default 3, up to 10)
{{recent_summaries}} / {{recent_summaries:N}}	One-line summaries of your latest entries (default 10, up to 30)
{{file:name.md}}	The contents of a Markdown file you pick via Insert Variable, or a path relative to your journal folder
{{today}}	Today’s local date, e.g. 2026-06-20 (Saturday)
{{now}}	The current local date and time in ISO 8601 format

Tip

Variables are resolved at the moment you submit the recording, so {{recent_entries}} reflects your journal as it was right then. Retries reuse the same resolved text.

Note

{{file:...}} can reference Markdown files outside your journal folder. Pick the file through Insert Variable and Ayumi keeps secure access to it. Files with the same name are kept separate, so a recent pick is never shadowed by another file.

Transcription Queue

Transcription tasks run in the background. You can:

• View the status of pending, running, and completed tasks
• Retry failed transcription tasks
• Delete tasks from the queue

On iOS, tasks continue processing in the background even when the app is not in the foreground. When a transcription finishes, Ayumi posts a Transcription Complete notification — tap it to open the resulting entry directly.

Transcription Results

Each Gemini transcription produces:

• Summary — A brief overview of the content
• Transcript — The full text transcription
• Response — The AI’s response based on your selected prompt
• Language — The detected language of the audio

On-device transcription, Gemini Transcribe, Muse Voice Transcribe, xAI Grok Speech to Text, and Groq Whisper produce the transcript only (plus the language where the engine reports it), since there is no prompt to respond to.

The resulting entry also records which model transcribed it and which profile was used. Both are shown in the entry detail view under a Transcribed label and stored in the Markdown frontmatter — see Markdown Format.

iOS Shortcuts Integration

See iOS Shortcuts for recommended Shortcut workflows, including background recording, Action Button triggers, and using Apple’s native transcription.