Read a voice message without uploading it

Short answer: Sendant can transcribe a voice message you receive on Android and iPhone. The optional 181 MB English/Russian model downloads once; after that, transcription runs on the device and works offline. The audio and the transcript are not sent to a server. Transcripts are encrypted at rest, excluded from linked-device sync and backups, and deleted with the message they came from.

What happens when you tap Transcribe

There is no background service and no automatic pass over your chats. You choose a voice message, and the app does the rest locally.

1. Install the model, onceOptional

   The first tap — or Settings → Offline transcription — offers a one-time 181 MB download of the English/Russian model. It is Wi-Fi by default; a cellular download needs its own confirmation. The file is digest-pinned, so an installed app only trusts the exact model it was built for. If you never install it, transcription is simply unavailable — there is no quiet fallback to a cloud service.

2. Decrypt the audio on the deviceLocal

   A received voice message is already on your phone as an encrypted file. Sendant decrypts it into a temporary working file, never into the player's shared temp directory, and never uploads it. For a long note delivered through the encrypted media path, the existing authenticated download is used first; if the media is unavailable, the app says so instead of transcribing something else.

3. Run inference on the phoneOffline

   A linked whisper.cpp engine reads the audio and writes text. The language is detected automatically between English and Russian, and you can switch it if the guess is wrong. Processing happens on the device's CPU: no speech-to-text API, no server, and no network needed after the model is installed.

4. Cache the transcript encryptedYours

   The result is written to an encrypted transcript cache with the same at-rest key as your message history. It is excluded from linked-device sync and from backup exports, and it inherits the source message's lifetime: delete the voice note, let it expire, clear the chat, or choose Clear transcripts, and the text goes with it. A late result cannot recreate a cache after the parent is gone.

What leaves the device

This is the part most transcription features are vague about, so here is the whole list.

Where it works, and where it does not

• Android and iPhone. The browser client does not transcribe; it has no local inference engine and Sendant will not send audio to a server to fake it.
• Incoming voice messages. The Transcribe action appears under messages you received. It does not transcribe your own recordings or run over a whole chat automatically.
• English and Russian. The pinned model covers both, detects the language automatically, and lets you choose explicitly if it guesses wrong. Other languages are not claimed.
• Offline after the download. With the model installed, transcription completes with every network connection blocked. The first use needs the download; later uses do not need a network.
• One job at a time. A second tap asks you to cancel the running job rather than starting a queue that competes with playback or messaging.

The honest limits

On-device recognition is a young feature, and the app says so in the transcript itself: “Transcribed on this device · May contain errors.” Accuracy varies with language, accent, background noise, microphone quality, and how clearly the speaker talks. The model is a compact multilingual package chosen to fit a phone, not a server-class model. Long notes take longer, and the first transcription after installing waits for the download.

It is also not a transcript archive. A transcript is a view of one message, not a new document with its own retention. If the voice note disappears, the transcript disappears; if you clear the chat, it goes too. That is the point: transcription should not quietly extend the life of a message you meant to disappear.