Emotion
Hear the cue, answer in a tone that matches it.
Conversational speech
Voice AI that keeps the feeling in the room.
Most assistants sound flat after the first minute. Vocalitic’s Conversational Speech Model is built to choose a tone, a pause, and a pace that fit the conversation, not just the sentence.
Four things that make speech feel human.
Timing
Pauses, interruptions, and emphasis learned from how people actually talk.
Context
The next line depends on what was just said, not on a fresh prompt every time.
A steady person
The same personality across a long conversation, so it can be trusted twice.
Where that voice is meant to work.
Care
- Companions that notice feeling
- Elder engagement
- Health information people can hear
- Mental health support
Work
- Customer conversations
- Meetings and transcription
- Spoken workflows
- Access to specialist tools
Home
- Assistants with a temperament
- Game and story characters
- Tutors
- Access tools
One model, not a stack of them.
Older speech systems split meaning and sound into separate models, and they ignore the conversation around the line. CSM is one transformer path: text and audio together, with the history of the talk used to pick the delivery. A smaller decoder reconstructs the sound so the reply can stay quick.
1M+ hrsMostly English audio, transcribed and diarized, in the published training set.
3 sizesTiny 1B, Small 3B, Medium 8B backbone. Decoders are 100M, 250M, and 300M.
~2 minContext window of 2048, enough for about two minutes of audio at a time.
More than the words
Emphasis, hesitation, pitch, and the other cues that change what a sentence means.
English first
Foreign names and mid-sentence switches are already in scope. Full support for more than 20 languages is the stated plan, not the current release.
More than one speaker
Distinct voices, turn-taking, and a tone that shifts with who is talking to whom.
What listeners actually preferred.
In a study on the Expresso set, eighty paid listeners compared CSM-Medium with human recordings. Without the prior conversation, they showed no clear preference. With about ninety seconds of context, they still preferred the human take. Isolated speech is at parity on that test. Fitting the delivery to the conversation is the open problem, and it is the one this work is aimed at.
The published business case.
$30B+
Addressable market by 2028, as the company states it.
12+ patents
Pending, in the portfolio described on this site.
How it would sell
Enterprise licenses, API access, and partnerships. A subscription margin is part of that description. Core research is described as headed for an Apache 2.0 release.