AI voices
AI voices for read-aloud in SovrGPT. Designed German voices with samples, nine languages and voices of your own, selectable in chat and via the Audio API.
SovrGPT reads answers aloud: in chat via the speaker button under every answer, and through the
API with POST /v1/audio/speech. Here you can listen to the AI voices in the catalogue and
find out how to design one of your own.
Designed, not recorded
The German catalogue voices are created from a written description (gender, age group, timbre,
pace) with the speech model Qwen3-TTS-VoiceDesign (Apache 2.0). They belong to no real person
and carry no biometric trait that could be attributed to anyone. The voice in the product tour
on the start page is bib-m-40.
The voices are spoken by Mistral at its EU endpoint (Privacy). Output in WAV, MP3, FLAC and Opus is labelled as AI-generated in the file header. In chat the label is visible on the button (AI labelling).
Nine languages
All catalogue voices speak German, English, French, Spanish, Italian, Portuguese, Dutch, Hindi and Arabic. They read the text in the language it is written in; there is no language switch.
We check German and English: every sample is transcribed back and must return word for word,
with no click at the start of a sentence. In the other languages the provider says a German
colouring remains audible. For English text without that colouring there is the stock voice
paul-en.
The catalogue
The catalogue currently holds six designed German voices, four male and two female between their mid-30s and early 50s, plus one English stock voice. The list is live. Each card has two samples (German, English) with the same sentence.
Loading voices …
Choosing a voice
- In chat: under Settings → Voices. Your choice applies from the next read-aloud on, for you as a user in every organisation with voice selection.
- Via the API:
voice: "bib-m-40"inPOST /v1/audio/speech. The catalogue comes fromGET /v1/audio/voices, including whether your organisation may select (Audio API).
Without a choice, Automatic applies: Mistral speaks, and for languages Mistral does not
speak (such as Turkish, Ukrainian, Polish), Supertonic reads aloud. Next to the read-aloud
button you see which engine spoke. You can also pick one of the ten fixed Supertonic voices
F1–F5 and M1–M5 (31 languages).
Voice selection is included from the Starter plan or with the "Voices" org entitlement; talk to us. Otherwise the stock voice remains. The samples here are open to everyone.
Design your own voice
You can describe a voice of your own and have it designed. As there is no speaker, you need neither a recording nor consent.
Whether designing is switched on in your environment is shown by
GET /v1/audio/voices/design-options in the available field. If it says false, an order
returns 501 not_implemented; you can still browse the options.
What you order
You pick traits from a casting sheet, the same dimensions casting platforms use:
| Dimension | Examples |
|---|---|
| Timbre | warm and round · clear and bright · full and dark · slightly smoky |
| Character | attentive · businesslike · calm · witty · trustworthy |
| Pace | leisurely with pauses · medium · lively · brisk |
| Intonation | calm · lifts towards the end · animated · assertive |
| Register | rather low · middle · rather high (professionals give hertz) |
| Use case | assistant · explainer · audiobook · news · commercial · phone announcement |
Add gender and age (16–90 years). Whatever you leave open is drawn: two orders without traits give two different voices, and the same order gives the same voice twice. The free-text field takes everything else, from "slightly hoarse" to "fundamental around 110 hertz". You can also describe the voice in your own words ("calm, a bit older, like a narrator in an audiobook") and let us propose the traits.
No real people. "Sounds like [name of a public figure]" will not be carried out; the suggestion helper turns a name into audible traits and does not adopt it. Describe only what can be heard, not appearance or origin.
In the app or through the API
- In the app (Settings → Voices): order, listen to the sample, adopt or discard. Nothing is adopted that you have not heard.
- Through the API:
POST /v1/audio/voices/designs. Withauto_adopt: truethe whole path runs in one go; without the field the draft waits for your approval (Audio API).
The design runs on a rented GPU server in an EU data centre (and additionally in Iceland or Norway, EEA) and takes a few minutes after a pause. After that a sample of roughly 20 seconds is ready. Every draw is measured (fundamental frequency, quartiles, voiced windows) and discarded if it misses the ordered register; the measurements come back with the answer.
Who owns the voice
A voice you design belongs to your organisation. Only your members see and select it, and it does not appear in the public catalogue. Its core is the reference clip: it stays with us in the EU and, for read-aloud, is registered as a voice with Mistral (EU endpoint). If the provider changes, the voice moves along without a new design.
Each organisation may hold up to ten own voices and three designs running at once, from the Starter plan or with the "Voices" entitlement. Voices in the public catalogue are registered by us only.
AI models
Every SovrGPT AI model with model ID, context, licence and cold start. Language models, embeddings, voice and the Decision Engine in EU data centres.
Decision Engine: AI decisions via API
Decision Engine: AI routes tickets, mail and enquiries and flags what needs review. It returns probabilities in milliseconds, via API and in chat.