返回资讯中心
外部精选
软件工程
#开发者实时流

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

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 m…

Hacker News 最新tinkeringdao2026/11/10 · 7 天前30 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

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:

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 by PromptEcho 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 and MalayalamTodoCues for Malayalam) relative to the recording's start timestamp, passing verified T1/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 paced 120 MB/s (0 MB/s during active recording) with MAP_POPULATE disabled, /proc/meminfo (MemAvailable) safety floors, and immediate page-cache reclamation (madvise(MADV_PAGEOUT)) on unload.

Deep Dive: How Satori Works

  1. Instant Capture, Categories & Hardware Power Key

Zero-wait microphone capture: RecordingService opens AudioRecord (16 kHz mono PCM) the instant you tap Record, 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, and To-do—plus custom user-defined categories (Settings → Categories, such as Journal). Each category maps to a recording style (SIMPLE_LISTENING, SHORT_MEETING, or LONG_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 via Change type (which rewrites the notes in the new category's style while preserving your title and to-dos).

Spoken languages & mid-recording switching:

English and German out of the box via NVIDIA Parakeet TDT 0.6B v3 (~670 MB).

Malayalam (മലയാളം) via AI4Bharat IndicConformer 120M INT8 (~140 MB). When Language in Record as is enabled, you can switch between English or German and Malayalam mid-recording between two sentences.

Custom vocabulary (Words to recognize): Add names, acronyms, and domain terms (one per line) in Advanced settings to boost their recognition in Parakeet v3 during live capture and Transcribe again.

Double-press power key (QuickKeyActivity): Binds to Motorola's hardware double-tap power gesture (com.motorola.mykey), debouncing accidental bounce presses within 1,000 ms (QuickKeyRepeatFilter) and waking over the lock screen (Hold to stop, Hold to change for pocket safety).

  1. Live & Post-Stop English ↔ German Translation (Translate Mode)

Selecting Translate in Record as (REAL_TIME_TRANSLATION) transcribes spoken English or German and translates finished sentences into the target language configured under Settings → Translate into (English or German).

Automatically detects the spoken language (German into English, or Already 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).

  1. The 7-Stage Post-Stop Finishing Pipeline (RecordingFinisher)

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)"]

Loading

How Malayalam Recordings Get Accurate English Notes (:feature-reader)

Compact 1B language models (Gemma 3 1B) struggle to parse Malayalam script directly, while loading 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:

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

Explicit reminders (explicit), first-person promises (commitment), and recordings made in the To-do category go straight to your To do list on Home (TodoStatus.OPEN) and schedule an exact Android alarm (AndroidReminderAlarm, with Done and Snooze 1 hr notification actions).

Tasks asked of you by someone else (asked), implied tasks (inferred), and appointment items land in the N suggested review sheet (TodoSuggestionsSheet) where you can Keep, Dismiss, Move to ideas, or Keep all.

Ideas (IdeasScreen):

Wishes or suggestions mentioned in a recording ("there should be an option to...", "it would be nice if...") and recordings made in the Idea category are collected on the Ideas 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).

  1. Cross-Recording Mind Map (MindMapScreen)

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 by Core (pinned + top themes), Recent (last 7 days), or All, and batch-index older recordings via Map N older recordings.

  1. Audio Playback, Re-Transcription, Retention & Offline Places

Audio player & seek-to-line: Play/pause, 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, or Malayalam (AI4Bharat)).

Automatic AAC compression & retention (AudioRetentionJob): Configure Compress 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) and Delete older audio (After N days, removing audio while keeping transcripts, notes, to-dos, ideas, and mind-map links). Per-recording Keep 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 or Near Berlin) with zero network requests.

On-Device Model Lineup & Memory Governance

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.

Model Role Languages Size Runtime

Silero VAD (silero_vad.onnx) Voice activity detection (bundled in APK assets; tunable pause length, max phrase, and threshold) Language-agnostic 1.8 MB ONNX Runtime 1.28.2

Parakeet v3 (PARAKEET_V3) Default live streaming speech recognizer & Transcribe again (supports custom hotwords via bpe.vocab) English & German (multilingual TDT 0.6B v3 INT8) ~670 MB sherpa-onnx

Parakeet v2 (PARAKEET_V2_EN) Optional Better transcript after Stop & Transcribe again for lower English WER English ~660 MB sherpa-onnx

Cohere Transcribe (COHERE_TRANSCRIBE) Optional high-accuracy Better transcript after Stop & Transcribe again English & German ~2.9 GB sherpa-onnx

Malayalam (AI4Bharat) (INDICCONFORMER_ML) Live streaming, post-stop second pass, and Transcribe again for Malayalam speech (IndicConformer 120M INT8) Malayalam (ml) ~140 MB sherpa-onnx

Malayalam Reading Bridge (:feature-reader, :lib-reader-onnx)

Model Role Direction Size Runtime

IndicTrans2 Distilled 200M INT8 Internal post-stop reading stage (segment_readings) so downstream LLM steps read English instead of raw Malayalam script Malayalam → English (mal_Mlym → eng_Latn) ~242 MB ONNX Runtime 1.28.2 C API + SentencePiece

Local Language Models (:feature-llm, :lib-llama-android)

Model Setting Name Role Size Quantization

Gemma 3 1B IT Standard (default) Live notes, post-stop notes, auto-naming, to-dos & ideas, live/post-stop EN↔DE translation, and mind map ~1.0 GB Q8_0 GGUF

Gemma 3 1B IT Efficient Lower-memory alternative for live and post-stop tasks ~650 MB IQ4_NL GGUF

Qwen 3.5 0.8B Qwen 3.5 0.8B Compact alternative model in Advanced settings ~870 MB Q8_0 GGUF

Gemma 4 E2B IT Gemma 4 E2B Optional higher-capacity model dedicated to AI for finished recordings (post-stop notes, naming, to-dos, and mind map when free RAM permits) ~3.3 GB Q4_0 QAT GGUF

Cooperative Memory Pacing & Disk-Cached System Prompts

Paced Pre-Faulting (ModelLoadPacing): Custom patches on llama.cpp (0001-load-progress-bytes.patch, 000

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →