Voices that sound like home

Text to speech, speech to text and voice agents, built Gulf first and for the organisations that serve people every day: banks, governments, contact centres. It is part of the sōtem family, and every sōtem model runs only on Mondaily's own servers.

Through a pilot. Tell us about your service; we set the engine up with the voices and transcription your callers need, agree pilot pricing, and go live with a small group of real callers first.

sōtem Khaleeji, our own Gulf model, is in training on recordings from paid native speakers who consent, through sōtem Data Lab. No open Gulf voice model is cleared for commercial use, so we build our own. Gulf transcription is available today.

Open "How this clip was made" under the player. It names the model, its licence and exact revision, the reference voice, and when and where the clip was generated. It also shows what an independent open speech-recognition model heard when it listened to the clip, so you can compare it with the script.

Pilot pricing per organisation. We agree it with you before the pilot starts, based on your volumes and where the engine runs. Public plans will follow once the service is live.

Every sōtem model runs only on Mondaily's own servers. There is no third-party AI API, and voices and text never leave our infrastructure. If your regulator requires hosting in your country, say so in your pilot request and we will tell you plainly what we can do.

The sōtem model family: sōtem V1, Scribe, Echo, Dub, Sound, Muse, Design, Clean, Guard, Agents and Khaleeji. Most are built on open foundations with commercial licences (Apache-2.0, MIT, CC-BY-4.0), each pinned to an exact version; sōtem Khaleeji is our own, trained on sōtem Data Lab recordings. Every foundation, its author and its licence are on the Licences & credits page.

Only your own. The speaker records a consent sentence that names them; we check it is in the recording's transcript and store the consent before anything is made. We do not copy anyone else's voice, public figures included.

Every clip the engine makes for customers carries an inaudible AudioSeal watermark (sōtem Guard), and a detector reads it, so anyone can check a recording. The example clips on this page predate the watermark.

The transcription model is trained for Arabic dialects and English. We have not published accuracy figures yet. In a pilot we measure it on your own calls, in your own dialects, before you decide.

Yes. Speaker labels and word-level timestamps are built into sōtem Scribe.

sōtem Design, a new voice from a description that belongs to no real person, is in development. We ship it only on a model whose training data is clean.

Yes. Pilot partners call the same engine from their own systems: text to speech, speech to text and voices. Every response names the model and licence that produced it.

Create a sōtem account to open the Voice studio. The same account works for sōtem Data Lab.

The world languages in the list above. They are served at language level: regional accents of English, Spanish or Portuguese are not separate voices yet.