Voices
The synthetic voices for read-aloud — designed, multilingual, with samples. Selectable in the app and via the API.
SovrGPT reads answers aloud — in the chat via the speaker button under every
answer, via the API with POST /v1/audio/speech. Which voice speaks is your
choice. This page shows the catalogue and lets you listen.
What these voices are — and what they are not
The German voices in the catalogue are designed, not recorded: generated from a
written description (gender, age group, timbre, pace) with a speech model
(Qwen3-TTS-VoiceDesign, Apache 2.0). They belong to no real person — there is no
biometric trait that could be attributed to anyone. The voice speaking in the
product tour on the start page is one of them (bib-m-40).
At runtime they are spoken by a European speech-model provider at its EU endpoint (see Privacy). Every output is labelled as AI-generated — in the file header (ID3/LIST/VorbisComment) and visibly on the button in the chat (AI labelling).
Multilingual
All catalogue voices speak the nine languages of the speech model: German, English, French, Spanish, Italian, Portuguese, Dutch, Hindi and Arabic. They render the text in the language it is written in — there is no language switch.
To be honest about it: the German voices were designed for German. German
and English are measured by us (every sample is transcribed back and must return
word for word; we also check that there is no onset click at the start of the
sentence). In the other languages the provider states that a colouring of the
source language remains audible — for English text without a German colouring
there is the stock voice paul-en.
The catalogue
As of 2026-09-19 the catalogue holds six designed German voices — four male (mid-30s, early 40s, mid-40s, early 50s) and two female (mid-30s, early 40s) — plus one English stock voice. The list below is live: a newly registered voice appears here without delay. Each card carries two samples (German, English) with the same fixed sentence, so the voices can be compared.
Loading voices …
Choosing a voice
- In the chat: pick the voice under Settings → Voices; the speaker button uses it from the next read-aloud on. The choice is yours as a user, in every organisation that has 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. Details in the Audio API.
Voice selection is included from the Starter plan. An organisation below that can be granted the “Voices” entitlement — talk to us. Without either, the system voice remains; the samples here are for everyone.
New voices
New voices are designed by us and registered in the catalogue — with description, primary language and samples. You need neither a recording nor a consent: a designed voice has no speaker. If you need a particular register or character, describe it to us.
Models
Language LLMs, embeddings, re-rankers and voice (STT/TTS) — all EU-hosted.
Coding in the IDE
SovrGPT as a model backend for Kilo Code, Cline, Roo Code, Continue, OpenCode, Qwen Code, goose, Zed, Theia, OpenHands, Cursor and Aider — sovereign EU models right in your editor, with honest figures on latency, telemetry, permissions and limits.