Updated: 2026-07-24
Agent voice in Corpis is two independent directions. Recognition (voice to text): the chat has a microphone button, you speak your message and it turns into text. Speech (text to voice): the agent reads its replies aloud in a voice you tuned to its character. Both run locally on the server, with no cloud keys.
Create an agent. Open agent creation and fill in the name, personality and greeting — or describe what you need in words and the AI assistant will fill the form.
Open the "Voice" tab. In the agent settings go to the "Voice" tab — this is where its voice profile lives.
Pick a preset. Take a ready character: Neutral, Stern, Cute, Child, Cheerful, Business, Friendly, Calm or Robot. Or describe the voice in words to the tab assistant.
Fine-tune axes and effect. Optionally adjust pitch, rate, liveliness, energy and breathiness; switch on the "robot" or "radio" effect if the character needs it.
Listen and save. Press "Listen" to check the voice before saving. Happy with it — save the profile.
Turn on speech in chat. Open a chat with the agent and flip the speaker toggle in the composer — from now on the agent reads new replies aloud.
Try voice input. Press the microphone button next to the input field, speak your message and stop the recording — the speech is recognized and becomes text.
The chat composer has a microphone button next to the input field. Press it, speak, press again to stop — your speech is recognized and lands in the field as ordinary text you can read through and send.
The audio is prepared right in the browser and reaches the server already in the needed form — no extra apps or codecs to install. Recognition runs on a neural model on the server, with no cloud and no keys.
The agent workshop composer has a speaker toggle. It is off by default. Turn it on and each new agent reply is also spoken aloud, while the Pip body moves its lips in sync with the loudness of the sound.
There is deliberately no fallback to the browser robo-voice: if server speech is unavailable for some reason, the reply simply stays as text in the thread — the character is not broken by someone else's voice.
The voice character is set per agent on the "Voice" tab in its settings. Start from a ready preset: Neutral, Stern, Cute, Child, Cheerful, Business, Friendly, Calm or Robot.
From there you can fine-tune the preset with character axes: pitch (higher is younger and brighter, lower is bigger and sterner), rate, intonation liveliness, energy and breathiness. A post-effect — "robot" or "radio" — is a separate switch. The "Listen" button plays the chosen voice right away, before you save.
Microphone recognition and spoken replies work in the on-platform chat. Speech in the web widget and in Telegram is not wired up yet — it is a next step, not a current capability.
A "child" voice on both engines is an approximation via pitch, not a separate child voice: the engines have no native child voices. A voice "by description or by a specific person's sample" is a class of GPU models, a separate future tier.
No. Both speech recognition and spoken replies run on a neural model right on the server, with no cloud and no keys. A key is only needed if you separately connect the cloud Yandex engine for more pronounced emotions.
Not yet. Microphone recognition and spoken replies work in the on-platform chat; widget and Telegram support is a next step, not a current capability.
Yes — there are ready "Child" and "Stern" presets plus fine axis tuning. Honestly: "child" is an approximation via pitch, the engines have no native child voice, and truly pronounced emotions come from the cloud Yandex engine.