How Walkie works

Where your voice goes.

Follow a dictation from microphone to finished text. See what stays on your computer, what goes to a provider, and what remains afterward.

walkie

VOICE → TEXT

01 / INPUTYour voiceMicrophone capture

02 / TRANSCRIBEParakeet / WhisperRuns on your computer

03 / OPTIONAL POLISHAI PolishLocal Gemma 4 E4B

04 / OUTPUTFinished textPasted into your app

Your speech stays on this computer.ON-DEVICE MODE

See every step

Desktop dictation · reviewed September 2026

01 / THE TWO ROUTES

Same shortcut. Different boundary.

The mode you choose decides where the speech-to-text work runs. The paths below describe desktop dictation; other features have their own routes further down this page.

A / PRIVATE PATH

On-device

NO SPEECH UPLOAD

Recognition and optional cleanup happen on your computer, so dictation works without a cloud speech service.

CaptureMICROPHONE → LOCAL AUDIO
You hold the shortcut and speak.
Walkie captures microphone audio on your computer. Silero VAD detects pauses before the recording is passed to speech recognition.
RecognizePARAKEET / WHISPER
A downloaded model makes the transcript.
The selected Parakeet or Whisper model runs locally using your CPU or GPU. Your audio is not sent to Walkie or a speech provider.
Clean up · optionalGEMMA 4 E4B / LLAMA.CPP
Gemma can polish the words locally.
If AI Polish is on and the optional intelligence pack is installed and ready, Gemma 4 E4B formats the transcript through a loopback-only llama.cpp process. Otherwise Walkie uses the raw transcript.
DeliverPASTE → LOCAL HISTORY
Text goes into the app you chose.
Walkie applies local rules for things like spoken numbers, pastes the result, and keeps transcription history on this device.

Speech and document context are processed on this computer. On-device dictation works without signing in; model downloads and optional signed-in features can still connect.

B / CONNECTED PATH

Cloud

AUDIO LEAVES DEVICE

Recognition runs through Groq when you choose cloud processing instead of a downloaded speech model.

CaptureMICROPHONE → LOCAL AUDIO
The same recording starts locally.
Walkie records your speech and uses local voice activity detection. A signed-in account is required for cloud dictation.
Upload boundaryDEVICE → WALKIE WORKER
The audio clip leaves your computer.
Walkie sends one dictation as FLAC, or WAV if FLAC encoding fails, to its Cloudflare Worker. The Worker checks your account and quota before routing to Groq. When AI Polish is on, it can also receive a destination-app category, selected text, up to 500 characters on either side of the cursor, or the sentence being rewritten.
RecognizeGROQ / WHISPER
Groq runs Whisper large-v3-turbo.
The Worker sends the clip to Groq. If a language check suggests Turbo translated speech into English, it retries with the full Whisper large-v3 model.
Clean up · optionalGROQ / GPT-OSS-20B
Groq can format the transcript.
When AI Polish is on, the transcript goes to Groq's openai/gpt-oss-20b and can include selected or nearby document text. With Polish off, Walkie sends no document context and returns the transcription alone.
Return & syncPASTE → ACCOUNT HISTORY
The final text comes back to Walkie.
Walkie pastes it into your app. Walkie does not store the dictation clip after processing; transcription history and related account data sync through Supabase in Cloud Mode.

Walkie does not retain the clip after dictation; Cloud Mode history does sync the resulting text to your account.

02 / MODEL REGISTER

Names, not mystery boxes.

These are the models configured for desktop dictation. Your local speech model is selectable; AI Polish is optional.

JOB	MODEL	RUNS AT	WHEN
Pause detection	Silero VAD v4	On your device	Both modes
Local speech-to-text	Parakeet TDT 0.6B v3 (recommended); v2; Whisper Small, Medium, Turbo or Large	On your device	Selected model
Local AI Polish	Gemma 4 E4B · QAT Q4 GGUF via llama.cpp	On your device	Optional pack
Cloud speech-to-text	Whisper large-v3-turbo; large-v3 on suspected mistranslation	Groq	Cloud Mode
Cloud AI Polish	openai/gpt-oss-20b	Groq	Optional switch

Model configuration can change with a release. This snapshot reflects the desktop app and Worker configuration reviewed September 2026.

03 / IMPORTANT EDGES

The details that belong in the diagram.

Privacy depends on where content goes and what remains afterward. Here are the parts a simple “local versus cloud” badge cannot show.

On-device speech is local. The whole app is not always offline.

Model downloads contact Walkie infrastructure without sending speech. Signed-in installs also check app status and can sync settings; opt-in analytics can make separate requests. An unsigned on-device install does not make the app-status check. If the signed-in status check stays unreachable for 90 days, the current desktop release stops working.

A failed dictation can leave a local audio file.

For recordings of at least 10 seconds, Walkie saves a recovery WAV before transcription so a failure does not erase your words. It is an ordinary file on your computer, cleared after text lands. A failed take remains until retry, discard, or a new long recording replaces it; there is no automatic expiry.

Cloud may process a clip more than once.

The Worker can start a second Groq attempt if the first stalls, and can retry on full Whisper large-v3 when Turbo appears to translate speech. Cloud failure can fall back to an installed local model.

Audio retention and text sync are separate.

Walkie does not store cloud dictation clips after processing. Cloud Mode transcript history, snippets, and dictionary data sync to your Supabase-backed account so they are available across devices.

Verbose diagnostics can contain your words.

Routine dictation logs are designed to record events and lengths. At Trace level, cloud requests ask for transcript fragments in response headers and write them to the local log. Older app versions logged more. A log stays on your computer until you export it; Settings → Support can scrub older words.

Turning off analytics does not stop usage accounting.

The analytics switch controls optional product events. Cloud requests still require an account and record usage counts for quotas and billing. The cloud infrastructure can also process request metadata such as IP address and time.

DATA LIFETIME

What remains afterward?

Deleting an account record, retaining request metadata, and keeping a recoverable backup are different things. Here is what we can confirm today.

Cloud dictation audio

Walkie does not keep the clip after processing. Groq's Inference APIs Zero Data Retention switch is enabled for eligible inference content.

Synced text and usage

Synced transcripts remain until you delete an entry or your account. Account-linked usage records remain until account deletion. Deleting them removes the active account records.

Local recovery audio

A failed recording of at least 10 seconds has no automatic expiry. Retry it, discard it, or replace it with a new long recording.

Meetings

Recordings and transcripts remain until deleted. A meeting shared with a team can remain in that team's archive after its recorder deletes their account.

Provider metadata

Groq retains non-content usage metadata. Cloudflare and Supabase can also process request metadata under their own schedules.

Project backups

The production Supabase dashboard reports a Free Plan, which has no scheduled project backups. We have not verified how long Supabase may retain any internal operational copies.

04 / WHERE OUR PROMISE ENDS

Cloud processing means another company sees the content.

Transport encryption protects a request while it travels. It does not hide the clip or the optional polish context from the provider that processes it.

WALKIE CONTROLS

What we keep

Walkie does not save the cloud dictation clip after processing. It does save the resulting text to Cloud Mode history and records account-linked usage. Meetings are different: their recordings and transcripts stay in your account until deleted.

PROVIDER CONTROLS

Groq inference ZDR is enabled

We have enabled Groq's Inference APIs Zero Data Retention for eligible requests. Groq says this prevents retention of request inputs and outputs for reliability and abuse review, while non-content usage metadata remains. Global ZDR is disabled. This setting does not cover Walkie's synced text, Cloudflare, Supabase, Soniox, or Speechify.

Read Groq’s data controls

Training: Walkie does not use your dictation to train models. Groq’s service agreement says it cannot train on inputs or outputs without the customer’s instruction. That is a contractual boundary, not a claim the desktop app can technically verify.

05 / OTHER SIGNAL PATHS

The mode badge is only one part of the story.

Read Aloud, meetings, mobile dictation, and agents handle different kinds of content. Their routes matter when you choose how to use Walkie.

READ ALOUD

Text becomes speech

The optional on-device voice uses Kokoro 82M. Cloud voices send the text you selected to Speechify, or to Soniox for supported voices. Read Aloud has its own voice-mode setting.

MEETINGS

Recordings have a different lifetime

Cloud meetings use Whisper large-v3 for transcription and GPT-OSS-120B for notes and chat. Meeting audio and transcripts are stored in your account until you delete them. On-device meeting processing stays local.

MOBILE

The keyboard takes another route

iOS Cloud Mode uses Groq; Android real-time dictation uses Soniox. iOS also offers on-device Apple recognition and Local Mode. These are separate from the desktop diagram above.

ASK AGENT

Your chosen agent has its own boundary

A local agent CLI runs under your account and sends questions wherever that agent is configured to send them. Walkie's hosted assistant is available in Cloud Mode only.

YOUR CHOICE / YOUR WORDS

Choose the boundary you trust.

For a detailed account of storage, analytics, logs, and deletion, read the privacy policy and the feature-by-feature data guide.

See what leaves your device Read the privacy policy