Manual · Reference

Privacy and data

What leaves your PC and where it goes, how to keep everything on your PC, where your files live, how long they are kept and how to delete them.

Last updated for A9 Voice 0.7.8

A9 Voice keeps your data on your PC. It calls speech and language services with your own keys, and only the ones you set up. There is no A9 Voice server in between, no account and no sign-in. This page lists every place data can go. Where something is a promise about how the app is built rather than a setting you can change, it says so.

The short version#

• Audio and text go only to the services you choose, with your keys. If you use models on this PC, nothing you say leaves it.
• One anonymous count, which you can turn off. The update check carries a random ID, the version, your channel, your Windows version and your country, so we can count installs, and five one-time messages (setup finished, first dictation, first note, Notes trial started, Notes bought) say how far people get. No crash reports, never what you said or recorded. See Anonymous usage counts.
• Your keys live in Windows Credential Manager and nowhere else: not in settings, logs, the repository or the app's own storage.
• Logs hold timings, lengths and status codes, never your words or your keys. This is also a rule of how the app is built.
• Voiceprints stay on this PC, are opt-in, and you can delete any of them or all of them at any time.
• Notes are plain files in a folder you pick.
• Recording consent is up to you. Some places require everyone's permission to record a conversation.

What leaves your PC, feature by feature#

The table says what is sent, to whom, and what happens when you use a model on this PC instead.

A few details that matter:

• Fallback. If the service you use for dictation fails or is slow, A9 Voice sends the same audio to your fallback service. If you do not want audio to reach a second company, set Use as fallback to a model on this PC, or to none.
• Summaries try more than one model. Summaries, Ask and Deep answers try the model you chose first and, if it fails or is rate limited, the next language model you have set up. If you want text to go to only one place, set up only one language model.
• OpenRouter is a routing service: your text goes to OpenRouter, which passes it to the model you picked. The same is true of any custom server you point at a remote service.
• A first Deepgram key asks before meeting audio goes there. When you save your first Deepgram key while Meeting notes use a model that can't tell speakers apart, the key window shows Also use Deepgram for Meeting notes, checked. Meeting audio goes to Deepgram only if you leave it checked. Unchecked, the key is saved and Meeting notes stay where they were. The setup step asks the same way and moves nothing until you click Continue.
• Deepgram live. With Live transcript on, a Deepgram key saved, and a Notes speech model that is not on this PC, both sides of the call stream to Deepgram live, whatever the Notes speech model is. To stop that, turn Live transcript off, use a Notes speech model on this PC, or disconnect the Deepgram key.
• Keeping the dictation connection ready. While A9 Voice runs and a cloud service is set for dictation, it sends that service a small request every few minutes, so the next dictation starts without delay. The request carries your key and no audio or text. With a model on this PC set for dictation, there is no such request.
• Dictation uses the clipboard briefly. It puts the text on the clipboard to paste it, keeps it out of Windows clipboard history and the cloud clipboard, and puts back what you had copied. When you ask for a copy yourself (tray Copy last transcription), that copy is allowed into clipboard history.

What each service receives#

The services A9 Voice can call are the ones in Settings › Models and Settings › Models. Each one only receives data when you set up its key (or address) and use it.

• Speech: Groq, Deepgram, OpenAI, ElevenLabs, and any server you add with Add custom API.
• Language models: OpenRouter, OpenAI, Anthropic, Groq, Google, and any server you add with Add custom API.

Each service's own terms say what it does with data it receives, and you pay it directly. A9 Voice has no copy of what you send.

Keeping everything on this PC#

To see the choice between a cloud service and a model on your PC:

0:43Speech models and your graphics cardWatch the 43-second video

You can use A9 Voice with no cloud service at all:

1. In Settings › Models, download a model under On this PC (Parakeet V2 for English, or a multilingual model). Under What each job uses, click Change on Dictation and Meeting notes and pick it. Set Fallback for dictation to another model on this PC, or none.
2. In the same page, Disconnect any cloud speech keys you no longer want. With no key saved, nothing can be sent there.
3. For summaries, Enhance and Ask, run a language-model server on your own PC and add it in Settings › Models with Add custom API. The page shows http://localhost:11434/v1 as an example address. Under What each job uses, click Change on Summaries and pick it; Ask and Enhance follow it. Disconnect any cloud language models, because summaries fall back to them.
4. Leave Settings › Integrations empty, and leave Connector for AI apps off if you do not want notes to reach another app.
5. For Live transcript, use a speech model on this PC that keeps up (its window says whether it does), or turn Live transcript off.

What still uses the network in this setup:

• The update check. A9 Voice checks a9voice.com and the project's release files on GitHub a few seconds after it starts and then every 6 hours. There is no switch to turn the check off; the anonymous count that rides on it has one (Settings › Privacy), and with it off the check sends nothing extra. If the PC is offline the check simply fails and tries again later.
• The Notes trial and license. Notes asks a9voice.com about its trial or key when you open it, press Try Notes free or buy (while you finish paying it asks every few seconds, for up to 15 minutes, whether the payment has gone through), a few seconds after A9 Voice starts and then about once a day while it runs (on the update check's timer), and when a recording starts and the last answer is a day old. A refunded or revoked key stops working within a day of the next check. When it can't reach a9voice.com it holds for 14 days (a trial) or 30 days (a key). Nothing waits for it: a recording in progress is never cut. Transcribe never does.
• Downloads you start. Speech models come from Hugging Face. Graphics-card support (about 25 MB, once) comes from files.pythonhosted.org. The speaker model for voice recognition comes from the project's release page on GitHub. All are fixed to one version and checked against a known checksum before use. After they are downloaded you can work with no internet at all.
• Web pages you open. The Get your key buttons open the provider's own page in your browser. That is your browser, not A9 Voice.

Anonymous usage counts#

A9 Voice sends no crash reports and no content-based analytics. The one thing it counts is installs and how far they get: with the update check it sends a random ID, the app version, the channel, your Windows version and processor type, and a9voice.com adds the country; and once each, with the same fields, a message when you finish setup, make your first dictation, record your first note, start the Notes trial and buy Notes. Installers downloaded from a9voice.com carry a random code in the file name. The app sends it once with the first-launch count so we know which link the download came from. It isn't tied to you. It is on by default and has a switch in Settings › Privacy. With it off, nothing extra is sent and nothing is recorded. The page Anonymous usage counts says exactly what is counted and what never is. The only network calls are to the speech and language services you set up, to the update locations above, to a9voice.com for the Notes trial and license (see the table), to the download locations above, and to webhook addresses you add.

Where your files live#

To see the folders and what Export offers:

0:45Your notes, your filesWatch the 45-second video

A copy of a recording is also kept briefly while it is being recorded and written up, in a temporary notes-spool folder in the app's data folder. It is removed afterwards.

Settings, the databases and the logs stay in the app's data folder on purpose: Documents may be synced by a cloud-storage service you use, and a synced database file can be damaged. The notes and audio folders are in Documents by default, so if your Documents folder is synced, they are too. That is your choice of folder, not something A9 Voice does.

If a hard crash damages the settings file or a database, A9 Voice keeps the damaged file beside the new one and tells you with a notice in the pill. Settings come back from the last good copy (or start from the defaults if there is none); a damaged history or notes database is rebuilt with the rows that could be read.

Your keys#

API keys are stored in Windows Credential Manager as generic credentials, on this PC. A9 Voice never writes them to settings.json, the logs, the exported settings zip or the app's own storage. The windows in the app only ever learn whether a key is set. Webhook addresses are stored the same way, because many contain a secret. Disconnect on a provider removes its key.

Copy diagnostics removes anything that looks like a key before it copies the last 200 log lines.

What the logs hold#

The log files record timings, lengths ("90 characters"), status codes, provider and model names, and the name of the app a dictation went into. They never hold API keys or what you said. The files are kept to about 2 MB each, with a few older ones; the oldest are removed.

One more file is local and worth knowing about: meeting-titles.jsonl in the logs folder. When Name notes after the meeting is on (or logging), A9 Voice reads the call app's window titles for the first 90 seconds of a call to find the meeting's name. This file keeps the titles it saw and the name it chose, for the last 200 calls, so you can check a wrong pick. It stays on your PC.

Voiceprints#

Recognize people by voice is off until you turn it on, after an explainer. When on, A9 Voice keeps a voiceprint for a person you tag in the Live window or name in a note: numbers that describe how they sound, not a recording. Voiceprints:

• are stored only in voices.db in the app's data folder, and the database is set up to overwrite deleted data;
• are never in your notes and never sent anywhere, including to the speech and language services, webhooks and the connector;
• are learned only from names you give, never from a guess;
• are deleted for good with Forget (one person) or Delete all voices in Settings › Voices.

Turning the feature on downloads the speaker model once. See Voices.

How long things are kept#

Exporting and importing your setup#

Export… in Settings › Account writes a zip with your settings, Dictionary, templates (with logos) and tag names. It does not contain API keys, webhook addresses or secrets, your history, notes, audio or voiceprints. The settings do contain folder paths, custom server addresses and your profile, so treat the file as personal. Import… reads such a zip and keeps what belongs to this PC (devices, folders, window places, setup and update state, and the voice-recognition switch). Keys are not carried over; add them again on the new PC.

The connector and webhooks#

Both let something outside A9 Voice receive your notes, and both are off until you set them up.

• The connector for AI apps lets an AI app on this PC search and read your notes. It is read only. It can see titles, summaries, transcripts, typed notes, image captions and action items (and your dictation history if Search dictations too is on). It never gets audio, images or voiceprints. Each request is logged on the Connectors page. See Connector for AI apps.
• Webhooks send a note to an address you add, when it is ready. The receiver gets exactly the parts you ticked, nothing else. Audio, image files and voiceprints are never sent. Whoever can read that service can read what you send, so add only places you trust. See Webhooks.

Recording consent#

Some places require everyone's permission before you record a conversation. A9 Voice records what your microphone hears and what your computer plays, and no bot joins the call. Getting permission is your responsibility, not the app's.

Delete your data#

Pick what you want gone. Everything below is on your PC only.

1. Dictation history: open Transcribe and use Clear all, or delete single dictations. Set Save audio recordings off in Settings › Dictation audio to stop keeping new audio.
2. Notes: in Notes, open the ⋯ menu of a note and choose Delete…, then Delete audio, files and note.
3. Voiceprints: Settings › Voices › Delete all voices.
4. Keys: Disconnect on each provider in Speech models and Language models. This removes the key from Credential Manager.
5. Webhooks: Delete… on each one in Integrations removes its address and secrets.
6. Models: open a model under On this PC and choose Delete.
7. Everything else: uninstall A9 Voice in Windows Settings › Apps. To also remove the files it leaves, delete the app's data folder, %APPDATA%\com.riptide9.scribe, the models folder in %LOCALAPPDATA%\A9 Voice, and your notes and audio folders in Documents\A9 Voice. Then open Windows Credential Manager, choose Windows Credentials, and remove any entries whose names start with com.riptide9.scribe/.

Related#

• Settings reference for every option.
• Speech models and Language models.
• Troubleshooting.

Something missing or wrong? Report an issue, or use Copy diagnostics in Settings › General when asking for help.