Satori – Offline Android voice notes with on-device STT and local LLMs

Fully offline voice recorder, live transcriber, journal & meeting synthesizer, and personal mind map for Android — powered 100% by on-device AI.

I wanted a way to capture thoughts on the go with a quick double-press of a physical hardware key—without unlocking my phone or getting pulled into a screen—and have those notes accumulate over time into a living mind map.

I use Satori every day for journaling, quick thoughts, ideas, self-development, local AI and engineering notes, and transcribing meetings. I didn't have the time to audition a dozen different note-taking apps, and I refused to pay recurring subscriptions just to hand my private voice recordings and personal reflections over to someone else's cloud servers. Instead, I built this.

Every stage of Satori runs 100% offline on the phone:

• Live speech recognition while you walk or sit in a meeting (NVIDIA Parakeet v3/v2 ,Cohere Transcribe ,AI4Bharat IndicConformer , andSilero VAD viasherpa-onnx and ONNX Runtime1.28.2 ).
• First-person note synthesis, session naming, to-do extraction, and mind-map linking via local instruction-tuned language models (Gemma 3 1B ,Gemma 4 E2B ,Qwen 3.5 0.8B ) running on custom-patchedllama.cpp ARM64 kernels.
• Neural reading & translation bridges for liveEnglish ↔ German conversation translation and an internalMalayalam → English reading pass (AI4Bharat IndicTrans2 200M ) so Malayalam voice notes receive the same structured English notes, reminders, and mind-map connections.
• Zero accounts, zero cloud servers, zero telemetry. The network is touched only when down model weights directly from Hugging Face.

I am sharing Satori as open source hoping it will be useful to anyone who wants to clone, customize, and build a private voice companion for themselves or for the people they love.

Built around the Grille typographic design system (Space Grotesk + Doto dot-matrix tabular counters + #D71921 recording red on #000000 OLED black), Satori is engineered for both standard Android displays (1080×2640) and the Motorola razr 70 plus external cover screen ( 1080×1272, 480×410 dp), including hardware power-key double-press capture over the lock screen and automatic layout clearance around the 232 px physical camera band.

Double-pressing the physical power key wakes the folded phone straight into recording (QuickKeyActivity), keeping all interactive controls clear of the 232 px hardware camera and flash band in both standard (ROTATION_0) and tent/flipped ( ROTATION_180) postures.

Each recording belongs to a category—Voice note, Meeting, Idea, To-do, or custom categories like Journal—that governs how the on-device LLM structures your notes and whether spoken action items become scheduled reminders or ideas.

Over days and weeks, Satori connects your recordings into a two-level knowledge graph (Themes and per-recording Topics) and gathers spoken wishes into a dedicated Ideas collection.

• 100% Offline & Private : Every byte of audio, every transcript segment, every synthesized note, and your entire mind map live in a local Room SQLite database (AppDatabase v13) on your phone. Nothing is ever uploaded to a server.

• "The Recorder Wins" Preemption : Starting a recording—from the Home screen, the cover screen, or a double-press of the physical power key over the lock screen—opens the microphone immediately and preempts any running background AI task within milliseconds.

• First-Person Notes, Zero AI Slop : Generic meeting bots write third-person filler ("The speaker discusses..." ). Satori's prompts and GBNF grammar schemas force the local LLM to write direct, first-person notes ("Pre-faulting GGUF weights at 120 MB/s prevents audio underruns" ), verified byPromptEcho filters.

• Calendar-Accurate To-Dos without LLM Math : Small language models hallucinate calendar arithmetic. Satori spots relative dates and times deterministically (TodoTimeResolver for English/German andMalayalamTodoCues for Malayalam) relative to the recording's start timestamp, passing verifiedT1 /T2 tokens to the model and grounding every extracted item against the cited segment (TodoGrounding ).

• Engineered for Real Phone Memory Limits : Multi-gigabyte GGUF models are pre-faulted at a paced120 MB/s (0 MB/s during active recording) withMAP_POPULATE disabled,/proc/meminfo (MemAvailable ) safety floors, and immediate page-cache reclamation (madvise(MADV_PAGEOUT) ) on unload.

• Zero-wait microphone capture :RecordingService opensAudioRecord (16 kHz mono PCM) the instant you tapRecord , buffering audio while the speech recognizer loads in the background so the first words you speak are never lost.

• Categories & Record as :

  • Ships with four starter categories—Voice note ,Meeting ,Idea , andTo-do —plus custom user-defined categories (Settings → Categories , such asJournal ). Each category maps to a recording style (SIMPLE_LISTENING ,SHORT_MEETING , orLONG_MEETING ) that governs how its notes and action items are structured.
  • Switch categories before recording, mid-recording via the top Record as chip, or retroactively on any finished recording viaChange type (which rewrites the notes in the new category's style while preserving your title and to-dos).

• Ships with four starter categories—

• Spoken languages & mid-recording switching :

  • English and German out of the box viaNVIDIA Parakeet TDT 0.6B v3 (~670 MB ).
  • Malayalam (മലയാളം ) viaAI4Bharat IndicConformer 120M INT8 (~140 MB ). WhenLanguage in Record as is enabled, you can switch betweenEnglish or German andMalayalam mid-recording between two sentences.

• Custom vocabulary (Words to recognize) : Add names, acronyms, and domain terms (one per line) inAdvanced settings to boost their recognition inParakeet v3 during live capture andTranscribe again .

• Double-press power key (QuickKeyActivity) : Binds to Motorola's hardware double-tap power gesture (com.motorola.mykey ), debouncing accidental bounce presses within1,000 ms (QuickKeyRepeatFilter ) and waking over the lock screen (Hold to stop ,Hold to change for pocket safety).

• Selecting Translate inRecord as (REAL_TIME_TRANSLATION ) transcribes spoken English or German and translates finished sentences into the target language configured underSettings → Translate into (English orGerman ).

• Automatically detects the spoken language (German into English , orAlready English: nothing to translate when input and target match).

• Displays a dedicated Translation tab, running either live every few sentences (As you speak ) or in one batch after stopping (After recording ).

When you press Stop, the microphone closes and RecordingFinisher processes finished recordings sequentially in LlmProcessingService while leaving the recorder free for your next thought:

flowchart LR
    Stop["Stop Recording"] --> Lines["1. Lines\n(Flush streaming STT)"]
    Lines --> Redo{"2. Better transcript\nafter Stop?"}
    Redo -- "Parakeet v2 / Cohere /\nIndicConformer" --> Retranscribe["Re-transcribe WAV\n(Making a better transcript)"]
    Redo -- "Off / Skipped" --> IsMl{"3. Malayalam\nrecording?"}
    Retranscribe --> IsMl
    IsMl -- "Yes" --> Reader["IndicTrans2 200M Reader\n(Malayalam -> segment_readings)"]
    IsMl -- "No (EN / DE)" --> Notes["4. Final Notes\n(Gemma 4 E2B or Gemma 3 1B)"]
    Reader --> Notes
    Notes --> Name["5. Auto-Name\n(2-5 word title)"]
    Name --> Todos["6. To-Dos & Ideas\n(Time resolver + LLM + Grounding)"]
    Todos --> Map["7. Mind Map\n(Topics -> Themes & Summaries)"]

Compact 1B language models (Gemma 3 1B) struggle to parse Malayalam script directly, while a 3.3 GB model (Gemma 4 E2B) on every voice note is too heavy when phone RAM is tight. Satori solves this with a dedicated internal reading stage:

• Visible transcript stays in Malayalam : Your recording'sTranscript tab displays the Malayalam script (മലയാളം ) produced byIndicConformer .

• Under-the-hood IndicTrans2 bridge : BetweenLines andNotes ,:feature-reader (:lib-reader-onnx ) runsAI4Bharat IndicTrans2 Distilled 200M INT8 (242 MB , 3 ONNX graphs on the bundled ONNX Runtime1.28.2 + SentencePiecemodel.SRC /model.TGT + IndicNLP normalization and Malayalam-to-Devanagari offset transliteration, beam size 5) to write internal English readings into Room (segment_readings , DB v13).

• Unified downstream synthesis :Notes ,Auto-naming ,To-dos & Ideas , andMind Map topics readsegment_readings , giving Malayalam recordings the same English notes, date resolution, and grounding quality as English recordings.

• Idle backfill (ReadingBackfill) : Malayalam recordings that do not yet have segment_readings are quietly upgraded when the phone is idle, without overwriting any title, note, or to-do edited by the owner.

• Open To-Dos vs. Suggested Inbox (TodoCapture) :

  • Explicit reminders (explicit ), first-person promises (commitment ), and recordings made in theTo-do category go straight to yourTo do list on Home (TodoStatus.OPEN ) and schedule an exact Android alarm (AndroidReminderAlarm , withDone andSnooze 1 hr notification actions).
  • Tasks asked of you by someone else (asked ), implied tasks (inferred ), andappointment items land in the**N suggested** review sheet ( TodoSuggestionsSheet ) where you canKeep ,Dismiss ,Move to ideas , orKeep all .

• Explicit reminders (

• Ideas (IdeasScreen) :

  • Wishes or suggestions mentioned in a recording ("there should be an option to..." ,"it would be nice if..." ) and recordings made in theIdea category are collected on theIdeas page (%d ideas, %d new on Home).
  • Ideas link back to their source recording and segment, outlast deleted recordings (Without a recording ), and convert into scheduled to-dos in one tap (Make it a to-do ).

• Wishes or suggestions mentioned in a recording (

• Two-level knowledge graph centered on You :

  • Themes (MindNodeKind.THEME) : Recurring subjects across your recordings (e.g.,Journaling ,Local LLMs & AI ,Deep Work ). When multiple recordings link to a theme,RefreshThemeSummariesUseCase writes a 2–4 sentence cross-recording summary of what is known and decided (newest recording wins when facts change).
  • Topics (MindNodeKind.TOPIC) : What an individual recording said about a theme (1–4 grounded words + up to 3 bullet points), linking directly back to the recording.

• Interactive Canvas & Accessible List : Toggle between the 2D radial graph (Show as map ) and a structured list (Show as list ), filter byCore (pinned + top themes),Recent (last 7 days), orAll , and batch-index older recordings viaMap N older recordings .

• Audio player & seek-to-line : , scrub bar, file size &Compressed badge, and tap any transcript line (Play from here ) to jump playback to that segment's timestamp.

• Transcribe again : Re-run speech recognition on any saved recording's audio with any installed model (Parakeet v3 ,Parakeet v2 ,Cohere Transcribe , orMalayalam (AI4Bharat) ).

• Automatic AAC compression & retention (AudioRetentionJob) : ConfigureCompress older audio ( After N days , shrinking~2 MB/min WAV files to~0.25 MB/min AAC/M4A while keeping them playable and re-transcribable) andDelete older audio (After N days , removing audio while keeping transcripts, notes, to-dos, ideas, and mind-map links). Per-recordingKeep this audio exempts important recordings.

• Offline GPS place naming (places.bin) : Optional Save the place resolves a recording's starting GPS fix offline against a bundled GeoNames KD-tree (app/src/main/assets/places.bin , e.g.,Berlin orNear Berlin ) with zero network requests.

Models are downloaded on demand inside the app (Settings → Models and storage and Settings → Advanced) and verified by byte size and SHA-256 ( ModelChecksums) before use.

• Paced Pre-Faulting (ModelLoadPacing) : Custom patches on llama.cpp (0001-load-progress-bytes.patch ,0002-map-populate-switch.patch ,0003-mapped-hook.patch ) disableMAP_POPULATE and pre-fault GGUF pages at a capped120 MB/s (0 MB/s while a recording is active) so disk reads never starve the UI thread or audio capture buffer.
• Page-Cache & Arena Reclamation : Releasing a model callsmadvise(MADV_PAGEOUT) /posix_fadvise(POSIX_FADV_DONTNEED) on mapped weight pages and purges native allocator arenas.
• Live /proc/meminfo Guards (MemoryFloor, AnswerFloor, ParseRoom) : Before Cohere Transcribe ,Gemma 4 E2B , orIndicTrans2 , Satori checksMemAvailable against model-specific headroom floors and falls back gracefully (e.g., fromGemma 4 E2B toGemma 3 1B ) when memory is tight.
• Disk-Cached System Prompts (forgetPromptCache) : All note, to-do, and mind-map system prompts are static and placeholder-free (strings_grille_ai.xml ,strings_grille_todo_ai.xml ,strings_grille_mindmap_ai.xml ), allowing:lib-llama-android to serialize their evaluated KV state to disk once and reload it across sessions without re-evaluating the system prompt on every recording.

Satori is organized as a 10-module Gradle project with strict Clean Architecture boundaries (:domain is pure Kotlin with zero Android dependencies):

flowchart TD
    app[":app\nCompose Grille UI, Cover Screen, QuickKey, Foreground Services"]
    presentation[":presentation\nViewModels, RecordingFinisher, HomeStrip, MindMap, Audio/Todo/Idea State"]
    domain[":domain\nUse Cases, Entities, Time Resolver, Malayalam Reading Prep, Grounding"]
    data[":data\nRoom DB (v13), Model Downloads, Memory Pacing, Offline Places, Retention"]
    feature_stt[":feature-stt\nStreaming & Offline STT + Silero VAD"]
    feature_llm[":feature-llm\nGGUF Inference, Prompt Cache & Grammar Constrained Output"]
    feature_reader[":feature-reader\nIndicTrans2 200M Malayalam-to-English Reader"]
    lib_sherpa[":lib-sherpa-onnx\nJNI Bridge to sherpa-onnx + ONNX Runtime 1.28.2"]
    lib_llama[":lib-llama-android\nCMake / JNI Bridge to llama.cpp (ARMv8.0 - ARMv9.2)"]
    lib_reader[":lib-reader-onnx\nCMake / JNI Bridge to ORT C API + SentencePiece"]

    app --> presentation
    app --> domain
    app --> data
    app --> feature_stt
    app --> feature_llm
    app --> feature_reader

    presentation --> domain
    data --> domain

    feature_stt --> domain
    feature_stt --> data
    feature_stt --> lib_sherpa

    feature_llm --> domain
    feature_llm --> data
    feature_llm --> lib_llama

    feature_reader --> domain
    feature_reader --> data
    feature_reader --> lib_reader
    lib_reader --> lib_sherpa

• JDK 17
• Android SDK : Platformandroid-35 (compile/target SDK 35,minSdk 29), Build-Tools34.0.0
• Android NDK :29.0.13113456 andCMake3.31.6+ (for:lib-llama-android and:lib-reader-onnx )
• Git submodules :third_party/llama.cpp andthird_party/sentencepiece must be initialized before building.

git submodule update --init --recursive

./gradlew testDebugUnitTest

./gradlew lintDebug

./gradlew assembleDebug

The compiled debug APK is written to app/build/outputs/apk/debug/app-debug.apk.

• GLOSSARY.md — Domain, hardware, and on-device inference terminology across the codebase.
• docs/GRILLE_DESIGN.md — Complete specification of theGrille visual system, color tokens, typography, and cover/inner screen layouts.
• docs/QUICK_KEY.md — Hardware power-key double-press (com.motorola.mykey ) integration and lock-screen capture behavior.
• docs/MODEL_UX_SPEC.md — First-run onboarding, background download state machine, storage checks, andHomeStrip nudges.
• docs/PRIVACY_AND_TERMS.md — Privacy architecture and local-only data storage guarantees.
• docs/THIRD_PARTY_NOTICES.md — Licenses and attributions for bundled libraries, fonts, GeoNames data, and downloadable models.