ThinkAloud

How ThinkAloud works

The whole pipeline, in plain language.

What it is

ThinkAloud is a note-taking app for people who think faster than they type. You talk — about an idea, a meeting, a recipe, a show you're watching — and the app turns it into a clean, organized Markdown note: a real title, topic tags, headings and bullets where they belong, checklists for the action items.

The unusual part is where that happens. The language model that does the writing runs on your own device's processor. Nothing you record is uploaded anywhere. Take the device offline and every feature on this page still works.

Capturing a thought

From the home screen you pick a note style — Journal, Meeting, Idea, Reference, or a template you wrote yourself — then hit the mic and talk naturally, "um"s and false starts included. You can also snap a photo (a whiteboard, a document, a slide) or type a quick line of text. The screen stays awake for the whole take, and the mic button becomes a stop button while you record — tap it and you can play the recording back before deciding: turn it into a note, or discard it. Nothing is processed until you say so.

A dedicated on-device speech model — NVIDIA's Parakeet — turns the audio into text fast and faithfully, keeping names, numbers, and dates, and labeling who said what in multi-voice recordings. The Gemma language model then writes the note, dropping the disfluencies; a second polish pass tidies the structure. You preview the note, edit it if you like, and save. If what you said overlaps a note you already have, ThinkAloud notices and offers to extend the existing note instead of creating a near-duplicate.

Long recordings

An hour-long lecture doesn't fit through a phone-sized model in one pass, so ThinkAloud processes it the way a person would: the speech model transcribes live while you record, the language model keeps a running understanding of the whole talk in overlapping segments, and at the end it writes one note in your selected style covering everything — followed by a timestamped transcript and the original audio, embedded and playable. On Android, processing keeps running in the background with a progress notification — you can use other apps while it works. On iOS the screen stays awake while ThinkAloud finishes in the foreground. Either way, you're notified the moment the note is ready.

Notes that organize themselves

Every saved note is plain Markdown with YAML frontmatter, stored in a folder called a vault — the same format Obsidian uses. You can point ThinkAloud at an existing Obsidian vault, pick any folder you like, or let the app manage its own. There is no database to export from; your notes are already files.

When a capture lands, the app searches your vault for related notes and adds a ## Related section of [[wikilinks]] — exact titles it verified, never links the model invented. Those links are what the knowledge graph draws: capture by capture, your notes wire themselves into a map of what you've been thinking about.

Asking your own notes

The AI Chat tab answers questions from your vault: "what did we decide about the rollout?", "where did I write down that sandwich recipe?". Retrieval finds the most relevant notes, the model answers from those excerpts only, and every grounded answer shows a Sources row — tap a chip and the note it came from opens. Follow-up questions work, voice questions work, and a deeper Think mode expands the search for harder questions.

What "on-device" actually means

On first run, ThinkAloud downloads two models from Hugging Face and stores them locally: Google's Gemma 4 E2B language model (about 3 GB) and NVIDIA's Parakeet speech-recognition model (about 0.7 GB) — those downloads are the only large transfers the app ever makes. From then on, transcription runs via sherpa-onnx, and writing, tagging, search, and chat run via LiteRT-LM on your device's GPU (or NPU on supported devices). A model picker in Settings shows alternative language models your specific device can run, and a toggle can hand transcription to Gemma instead — much slower, but it understands languages beyond the speech model's English.

Practical requirements: roughly 4 GB of free storage for the models and a reasonably recent device. Newer devices process faster; the feature set is identical.

Sync, if you want it

Sync is off by default. If you turn it on, your vault syncs to your own Google Drive (a private app folder) or iCloud — authenticated by you, on your account. ThinkAloud has no servers of its own, so there is no middleman to trust: the developer cannot read your notes even in principle. The same vault can be shared between your phone and the desktop app.

What it costs you

No subscription and no account — there's nothing to sign up for. The model is free; the app does the work with hardware you already own.

See every feature, with screenshots →