Privacy Policy

Last updated: 17 October 2026

Speech is a dictation app: it turns what you say into clean written text and puts it into whichever app you are using. This policy is what it collects, why, and what you get to decide. Speech is operated by Rupees Venture Private Limited ("we", "us"), Indore, Madhya Pradesh, India.

1. Information we collect

Beyond that pre-cursor snippet, the text you select in Agent Mode, and Screen Vision if you have switched it on, we do not collect your contacts, location, photos, or the contents of the screen or apps you use. The accessibility permission does two things and no more: it places your dictated text at the cursor in the field you are typing in, and reads that snippet (see Section 5).

We use no cookies for tracking, and this website carries no advertising or analytics code. When you sign in, your session token and your dictation key are held in your browser's local storage so you stay signed in; signing out removes them.

2. Microphone and recording

The microphone is used only while you are actively dictating, holding the dictation button or bubble. Speech does not record in the background and does not listen when you are not dictating, unless you deliberately start a meeting recording or switch on the optional wake word.

Meeting notes: starting a meeting recording captures your microphone and, with your Mac's permission, call audio. Audio is streamed for transcription and is not saved by the meeting service. On Mac version 0.9.43 and later Speech also saves a local recording with your microphone and call audio in separate channels, so you can play it back while reviewing your notes. That local recording is kept in both Standard and Privacy Mode. Text is processed online to produce the transcript, summary and action items. In Privacy Mode the notes are saved on your Mac and snapshots sent for processing are not saved to the cloud meeting archive; outside Privacy Mode, transcripts and notes live in your Speech account until you delete them. Recording stops when you press Stop or the meeting duration limit is reached. Tell the other participants before you record.

Local meeting audio: those recordings stay in Speech's application support folder on the Mac that made them until you remove the files. Speech does not upload that saved audio and does not delete it on a timer. Show file in a saved meeting finds it in Finder. Deleting a meeting transcript, signing out, or deleting your Speech account does not remove that separate local file; delete it in Finder when you no longer want it. Earlier meetings have no local recording unless audio was saved at the time.

3. How your data is used

• To turn your speech into text and give it back to you.
• To provide your dictation history inside the app and your account.
• To keep the service running, secure, faster and more accurate.

We do not sell your personal data, and we do not use your dictation content for advertising.

4. Third-party processing

To transcribe and format your speech, audio and text are handled by these providers, acting on our behalf:

• Sarvam AI: speech-to-text transcription. Receives your audio.
• OpenRouter: routes model requests for us. Depending on the path a dictation takes, the model that runs is operated by Google, which may receive the audio itself, or Groq, which receives the transcribed text.
• OpenAI: receives selected text when you use a text command or rewrite. Never audio.

They process it only to do that transcription and formatting for us. Three more providers keep the service running without touching what you dictate:

• Amazon Web Services: hosting. Our servers and databases run in AWS's Mumbai region (ap-south-1), in India.
• ZeptoMail (Zoho): sends your sign-in codes and any service email. Receives your email address, never your dictations.
• Google: if you choose "Continue with Google", Google confirms who you are and we receive your email address. Separate from Google's role as a model provider above.

If you connect an optional integration yourself, Swiggy for Agent Mode ordering for instance, we hold an encrypted access token so the feature can act for you. Disconnect it whenever you like; disconnecting deletes the token. These integrations stay off unless you connect them.

5. Accessibility permission (macOS, Android) and keyboard shortcut (Windows)

On macOS, Speech uses the Accessibility permission for one purpose: dictation. It puts your dictated text at the cursor in whichever app you are in, and reads the short snippet of text immediately before that cursor so spacing and capitalisation come out right. It does not read the rest of the document, does not read other apps, and collects nothing when you are not dictating. Speech also registers a system-wide shortcut so you can dictate without leaving the app you are in; like any global shortcut it sees key events across the system, but Speech acts only on the shortcut you chose and never logs, stores or transmits what you type.

On Android, Speech uses the Accessibility Service for one purpose: dictation. It notices when you focus a text field so the bubble can appear, inserts your text at the cursor, and reads a short snippet immediately before it so the result is spaced and formatted correctly. It does not read other apps or the rest of your screen, does not run when you are not dictating, and never appears in password or payment apps. With dictation inactive it collects nothing.

On Windows, Speech registers a system-wide keyboard shortcut so you can dictate without leaving the app you are in. Like any global shortcut the hook receives key events across the system, but Speech acts only on the one you chose, and does not log, store or transmit what you type. To place the finished text it copies it to your clipboard, sends a paste keystroke to the focused app, and then puts your previous clipboard contents back. It does not read the contents of other applications.

5a. Screen Vision (Agent Mode)

Agent Mode can answer questions about what is on your screen ("is page pe API key kahan hai"). For that it needs a picture of the window you are looking at, which makes it the one feature that captures your screen, and it is off by default. Nothing is captured until you turn Screen Vision on yourself and grant your operating system's screen recording permission.

With it on:

• Only the focused window is captured, not your whole screen, and only on a request that actually refers to the screen. A plain dictation never sends an image.
• Speech never captures a password manager, keychain, wallet or banking app, and never its own windows.
• The image is held in memory, sent to the model answering your question, and never written to disk on your device.
• Turn it off in Settings, or revoke the screen recording permission in your operating system, and capture stops.

5b. Agent Mode: the selection, and the short conversation memory

Agent Mode is the separate control that asks Speech to do something rather than type what you said. Two things reach our servers on those requests that a plain dictation never sends.

• The text you have selected. Highlight a sentence, say "isko formal banao", and that sentence is what we are being asked to rewrite, so it goes with the request. Nothing is sent when nothing is selected.
• A short conversation memory. Your last few Agent Mode requests and their replies are kept for five minutes so a follow-up means something. It is scoped to the app you are working in, is never used to build a profile of you, and expires on its own. Saying "bhool jao" deletes it immediately, as does deleting your account.

Agent Mode requests go to the same providers listed in Section 4 and carry the same retention choice from Section 6 as your dictations.

5c. Connecting your Google account

Google sign-in uses your verified email address to create or open your Speech account. It grants no access to Gmail, Calendar or Drive. Connecting Google is a separate, optional consent: it lets Meeting notes show your calendar events and, where Agent Mode is available, lets it work with your Gmail, Calendar, Sheets and Docs. You choose the account and the permissions.

• We hold your refresh token encrypted, tied only to your Speech account, and use short-lived access tokens to make requests. Disconnecting in Settings deletes the stored connection and revokes its Google grant. You can also revoke access from your Google account's security page.
• Calendar events are fetched to show your meetings. Agent Mode reads Google data only for the request you make, and passes the relevant part to the model answering it. A reply may sit in the short conversation memory described in Section 5b so you can ask a follow-up. Google data is not used to train models, is not used for advertising, and is not shared with other Speech users.
• Nothing is ever sent, created or changed in your Google account without a confirmation you tap first.

Speech's use and transfer of information received from Google APIs follows the Google API Services User Data Policy, including its Limited Use requirements.

6. Storage, security, retention, and your Privacy Mode choice

Your data travels over encrypted (HTTPS/TLS) connections and is stored on secured servers in India. How long we keep the content of your dictations is your choice, in the app under Settings → Privacy:

• Standard. Your transcripts are stored on our servers for as long as your account exists, and may be used to improve Speech's accuracy and models. They are never sold, and never used to train anybody else's models. The voice recording itself is kept for up to 60 days and then deleted automatically. You can delete everything whenever you like (Section 7).
• Privacy Mode. Your audio and transcripts are transcribed and then discarded immediately. Nothing is kept on our servers, and no dictation history is kept on your device.

Anonymous usage counts, the number of words dictated for instance, are kept in either mode and contain no transcript text and no audio. You can delete your account and everything with it at any time (see Section 7).

Pictures you asked Agent Mode to generate are kept for up to one year, along with the words you used to ask for them, because the app has a page that shows you your own pictures and a picture you cannot find again is one you have to pay to make twice. Deleting your account deletes them.

Bug reports are kept longer than dictations, for up to one year, including any screenshot attached. A screenshot attached to a bug is a record of a defect rather than a recording of your voice, and a year is about how long a report stays useful while the defect is being fixed. They are deleted automatically after that, and Section 7 says what happens to a report when you delete your account.

Crash reports are kept for up to 90 days and then deleted automatically. Privacy Mode does not change them, because a crash report carries none of what you dictate. They are not linked to your account, so there is nothing to find when you delete it; they expire on the same 90 days.

Nobody at Rupees can read your transcripts or play your audio unless you switch that on yourself, from your own account page, and no administrator can switch it on for you. Without it, what our own team can see about a dictation is when it happened, which app it went to, how long it was, how many words it produced, and whether it worked, which is what we need in order to fix things.

7. Your choices and rights

• Delete everything in the app. Settings → Account → "Delete account & data" permanently erases your account and every dictation, transcript and recording stored for it, from our servers and your device. It is immediate and irreversible.
  One exception: if you sent us a bug report, the report itself stays in our engineering queue with you removed from it. Your account link, your name and the device details attached to it are erased with everything else; what remains is the description of the problem, which is the part we needed in order to fix it.
• Or ask us. Email support@rupees.com from your account's email address and we will provide or delete your dictations and account data for you.
• Choose what we keep. Privacy Mode (Section 6) is a switch in the app and you can change it whenever you like. It applies from the moment you turn it on.
• Screen Vision is off until you turn it on, and turning it off stops any capture (Section 5a).
• Permissions. You can revoke microphone, notification, overlay, screen recording or accessibility permissions at any time in your operating system's settings, on macOS, Windows and Android alike. Some features stop working without them.

8. Children

Speech is meant for people aged 18 and older and is not directed at children.

9. Changes to this policy

We may update this policy from time to time. Material changes are reflected by the "Last updated" date above.

10. Contact and grievances

For any privacy request or question, write to support@rupees.com. We aim to reply within two working days.

If you are not satisfied with how we handled a request, you can escalate it to our Grievance Officer under the Digital Personal Data Protection Act, 2023, at grievance@rupees.com. We acknowledge a grievance within 72 hours and respond within 30 days.

Rupees Venture Private Limited, Indore, Madhya Pradesh, India.
See also our Terms of Service.