Terms Update: Our Own Mongolian Speech Model, and Opt-In Contribution

The short version

From October 1, 2026, Duudlaga Flow will begin training its own Mongolian automatic speech recognition (ASR) model.

What matters most:

• This is voluntary. The setting ships off for every user.
• If you never turn it on, nothing changes for you.
• If you do turn it on, only recordings made after that point are included.
• Nothing recorded before you opt in is ever used.

Why we are doing this

Today, Duudlaga Flow relies on foreign providers to turn Mongolian speech into text. Those models were never trained specifically for Mongolian, so they stumble on Mongolian names, domain terms, and regional pronunciation.

The only real fix is a model trained for Mongolian. Training one takes real, everyday Mongolian speech — not clean read-aloud recordings.

So we decided to do this by asking, not by taking.

What actually changes

Settings will gain a new option: "Contribute to the Mongolian speech model."

If you leave it OFF

Nothing changes. Your recordings are processed and discarded as before, your transcripts stay encrypted, and nothing enters training.

If you turn it ON

Recordings you make after that point, and their transcripts, are stored on our servers and used to train our Mongolian speech model.

We ask you — we do not take quietly

The setting is not buried somewhere for whoever happens to find it. We put the question to every person once, directly:

• When you register,
• At the end of onboarding,
• On your next sign-in if you already have an account and have not answered yet.

Three things worth stating plainly:

• Not answering is not agreement. Dismiss or postpone the question and no answer is recorded, and nothing is collected.
• Declining is respected. If you say no, we do not put the same question to you again.
• Only an explicit yes starts collection — and even then, only for recordings made after you answered.

Policy change history

Alongside this change we are opening a Policy change history page. Every material change to the Privacy Policy and the Terms of Service is logged there, with its date and the sections it touched.

The reason: a policy that can be edited quietly is not a commitment. You are entitled to know what the policy said at the time you agreed to it.

Why earlier content is off limits

Until now we promised we would not use your data to train AI. You agreed to our terms on the strength of that promise.

Updating our terms does not, in our view, give us the right to reach back and change what that promise covered. So:

• Content from before you opt in is never brought into training.
• There is no path in our systems that backfills older data into the training set.

There is also a practical reason: we have never stored raw dictation audio at all. It was discarded immediately after processing. That data simply does not exist.

The two copies differ — plainly stated

We think this part has to be said clearly.

Your history — the transcripts and notes in the app — stays encrypted under a key derived for you alone. Our team, the founder included, has no way to open it. That does not change.

A contributed copy is stored separately, and our training team can read it. There is no way to train a model on data nobody can open.

So contributing means agreeing to show those specific recordings to our training team. The app makes this explicit before you turn the setting on.

What we still will not do, even with contributed data

• Not sold or rented.
• Not given to third parties — training data stays inside Duudlaga Flow LLC.
• Not published or released as an open dataset.
• Not used for advertising.

You can take it back

• Turn the setting off — future collection stops immediately.
• Delete what you contributed — a separate action permanently removes the recordings and transcripts you already contributed. Deleted material leaves the training set.

As before, "Clear my data" and "Delete account" also remove your contributed recordings.

What is NOT changing

Worth stating outright, because it would be easy to misread:

• Speech-to-text is still performed by Google. Our own model is not ready, so transcription continues exactly as it does today.
• AI polishing and translation still send text to OpenAI, if you have those features enabled.
• Section 3 of our Privacy Policy, covering third-party processing, is unchanged.

Reducing our dependence on outside providers is the goal once our own model is good enough. That day has not arrived, and we are not going to describe it as though it has.

Timeline

If you never turn the setting on, October 1 changes nothing for you.

Questions

If anything here is unclear, write to [email protected].

Your voice. Your thoughts. Your words. Sharing them is your call.