Embeddings and Voice Mixing

Cartesia models represent voices in the form of embeddings. An embedding is a length-192 vector of floats between -1 and 1. Collectively, these numbers capture the characteristics of the voice: speed, emotion, accent, and so on.

Our text-to-speech endpoints currently require embeddings to be specified for each generation. Support for specifying voice IDs instead of embeddings is planned.

Voice Mixing [Alpha]

A neat feature that embeddings enable is voice mixing. You can interpolate two embeddings to obtain a third voice that sounds somewhere between both of them. This feature is available in the playground.

We suggest using linear interpolation to interpolate between embeddings. To interpolate between embeddings $A$ and $B$ to obtain $C$, use the following formula:

$C = (1-\alpha )A + \alpha B$

$\alpha$, or alpha, is the “interpolation coefficient”—an alpha of 1 means $C = B$ and an alpha of 0 means $C = A$. For example, with $\alpha=0.5$, $C=0.5A+0.5B.$

Prototype embeddings

Our playground’s voice design features (i.e. the speed and emotion controls) rely on prototype embeddings. These are embeddings that capture the essence of some characteristic: fastness, slowness, anger, curiosity, and so on. If you interpolate a voice with a prototype embedding, it will acquire the characteristics of the prototype embedding.