Wake word
Listen stays on after you leave. A watchdog restarts it if the microphone stalls, and Settings shows that status. Nothing starts until you say “Hey Grok.”
Android · on the phone by default
Say “Hey Grok” with the screen off. A new install answers on the phone, with no cloud key. Notes, calendar, reminders, and money stay on the phone. xAI and OpenAI are optional.
Arm64 APK · Android 8 and newer · signed in GitHub Actions
Listen stays on after you leave. A watchdog restarts it if the microphone stalls, and Settings shows that status. Nothing starts until you say “Hey Grok.”
After a pause you choose, from 5 to 120 seconds, the conversation ends and only the wake word remains.
Titles, search, and Markdown. Preview headings, lists, and links before you save.
0.1.44 makes expenses and incomes fully editable, including amounts. Coming expenses and Coming income are limited to the next 30 days. Credit cards are payment methods apart from accounts. A charge raises card debt and hits this month's result without moving the bank. Paying the card from a funding account is a debt payment, not an expense, and does not change net worth. Talk has voice parity for Finance: accounts, investments, cash, debts, cards, expenses, incomes, charges, and pay-card. A whole-app translation audit covers English, Spanish, German, Portuguese, and Valencian (en, es, de, pt, ca-ES-valencia), including calendar titles that were previously hardcoded. 0.1.43 makes Finance phone-first. Filters are a scrolling date-range row plus a Filters sheet, and the action bar sits above the Finance tabs so Program a bill is not covered. The same clipping pattern is tightened elsewhere. Live daily FX uses ordered multi-provider failover. api.frankfurter.app now 301s to api.frankfurter.dev, and the old redirect allowlist refused the new host, so every fiat fetch failed. Fiat now tries Frankfurter on api.frankfurter.dev, then open.er-api.com, ECB, and currency-api on jsDelivr and pages.dev. Crypto tries CoinGecko, Coinbase, Kraken, then currency-api. One book is kept per local day. Settings Refresh forces a fetch. Manual rates in Settings are the last resort. There are no invented rates, no 1:1, and no silent USD fallback. Unpaid scheduled expenses, payments, income, and investment payouts appear as all-day items on the in-app Calendar on each due date, for example rent on the 20th. Existing Time reminders and Spoken alerts are unchanged. 0.1.42 opens Finance on net worth, then this month's expenses, then income, then the month's result. A thumb-zone tab bar lists Overview, Expenses, Income, Accounts, and Charts. The rate book is only in Settings. Live Frankfurter and CoinGecko rates still load in the background as in 0.1.41. 0.1.41 uses today's live exchange rate, with manual rates in Settings. Finance fetches Frankfurter and CoinGecko when the screen opens, when the local day changes, and when a needed pair is missing. A rate fetched today wins, then a labeled stale cache, then a Settings manual rate unless that rate is marked as an override. Overview sections share one column width. 0.1.40 shows combined Finance totals in the Settings currency on Overview, charts, the PDF, and voice. A missing rate asks for a real rate instead of leading with a per-currency stack, and native amounts stay on each item. Themes add Midnight, Carbon, Ivory, Sand, Forest, Plum, and a custom palette. Home-screen widgets follow the theme and show those combined totals. 0.1.39 adds Finance schedules for one-off and recurring income and expenses, including fixed-term runs such as a 9-month 270 EUR course, open-ended rent with no lifetime total, subscriptions, planned income, and investment payouts such as Nitroeggs with a quoted value settled in another currency. Overview is five cards: net worth right now, this month's cash flow, Coming expenses, Coming income, and If these land. Display-only multi-currency conversion uses Frankfurter, CoinGecko, cached, or manual rates; native amounts stay, and a missing rate omits the combined total. Next 30 days is its own panel of planned in and out, with overdue listed separately. Missing Spanish, German, Portuguese, and Valencian UI, PDF, and voice strings are filled in, PDF accents are distinct, and phone layout fixes keep Finance panels from scrolling sideways, wrap the program sheet, and label reminders for one day and a week. 0.1.38 adds an app-wide magazine look, Graphite, Ink, Stone, and System themes, a keyboard inset under the focused input, a default currency in reorganized Settings, and Finance cash-flow, upcoming bills, and debt payoff. 0.1.37 adds separate Finance accounts, phone charts, a PDF summary, and payment reminders. 0.1.36 adds a default music player, Musify voice control, and searchable Settings. 0.1.35 adds an on-device Finance tab and Talk money tools, optional Spanish speech and voice packs, LFM2.5 350M for on-device Talk, local music from Talk, and a custom audio file or YouTube sound on in-app alarms. Spanish, German, Portuguese, and Valencian screen copy is complete, and the phone layout was reviewed for one-handed use. 0.1.34 fixes the blank screen after launch for already-onboarded users (the 0.1.33 hook crash). In 0.1.33, Settings lists all 28 Grok voices and the 10 OpenAI Realtime voices, each with a preview. System back walks in-app screens. Back and add sit in the bottom bar. A wake-word watchdog restarts listening if the microphone stalls, and Settings shows the background-listening status. Spoken mode posts the banner on a silent channel, so the normal chime no longer plays over the voice. If a sound still plays, speech waits for it. The spoken voice plays at full level on the alarm stream, and the phone’s alarm volume is the only ceiling. An optional name in Settings lets reminders be spoken in the second person, for example “Hey Gabriel, remember to brush your teeth.” Normal mode is unchanged and remains the default. Reminders, alarms, countdowns, and calendar events created in Talk stay inside Grok Voice and fire through this app’s notifications. That in-app schedule shipped in 0.1.30. Talk does not write those creates to the phone Clock. Person-date visibility is unchanged.
Settings has one Voice and AI screen. From there you pick who hears you, which model thinks, which voice speaks, and the order Talk tries if the first one fails. On this phone can sit next to xAI and OpenAI. A key is typed on the provider it belongs to. Downloaded, key set, and offline-ready show at the top.
The first launch downloads one English pack: Moonshine, SmolLM2, and Phonon 31m. It shows the size and the progress, and it asks for Wi-Fi. There is no Hugging Face token and no other-language pack in that tour. The screen language can still be Spanish, German, Portuguese, or Valencian. Those voices, Piper, and the gated Phonon models can be added later in Voice and AI. The tour can be paused and finished later from that screen. An install that already finished setup is not sent through it again.
The on-device engine is Phonon, Piper, or the Android voice, or Auto. Auto uses Phonon when that language already has a model, otherwise Piper, then the Android voice. If Phonon fails, Piper speaks, and the Android voice is next.
Phonon is Gradium’s on-device speech engine. It builds on Pocket TTS and the Mimi codec from Kyutai. The runtime is xn-ptts, MIT or Apache 2.0, and it runs on the phone with no PyTorch. English Phonon 31m is Apache 2.0 and needs no account. The 40m models are CC BY-NC 4.0 and gated on Hugging Face. This app is non-commercial, so those terms allow them. Accept the license on the model page, then save a Hugging Face token in Voice and AI. The token stays in the on-device vault and is sent only for that gated download. English 31m works with no token. Valencian has no Phonon model, so it keeps Piper when a voice is installed and otherwise the Android voice. Portuguese has no Piper voice; Phonon speaks it after the model is downloaded.
Moonshine speech-to-text, SmolLM2, and the Piper packs are still there, under the same screen. Nothing new downloads until you tap Download. After that, the voice is offline.
This app has no server of its own and does not collect analytics. The key is stored encrypted on the phone. Wake-word audio is checked on the device and is not uploaded. Notes, reminders, and in-app events stay in a local file.
The default conversation stays on the phone. If you add a cloud provider, that part of the talk is not private from them. An xAI key is sent to api.x.ai to mint a short-lived voice token, and speech goes to xAI only while xAI is the provider in use. OpenAI works the same way with its own key. A typed question that uses a cloud model is sent there too.
Download the APK, then install it from the phone or from a computer.
adb install -r GrokVoice-0.1.48.apk