Two models, one API key

Text to speech and speech to text, one key, one balance, nothing to wire up between them

Tokay-1.0 · text to speech

Trained on speechyou can’t scrape

Tokay is trained end to end on voice data we record ourselves. That is why it speaks languages the incumbents can’t, and why the quality holds where scraped models fall apart.

It sounds like a person, not a text-to-speech voice, and it starts talking fast enough that nobody waits for the first word.

What it speaks todayThe voice catalogue →
Languages
Kazakh
Voices
Six, three female and three male
Model
tokay-kk-v1
Next
More languages on the same endpoint
Seta-1.0 · speech to text

It never askswhich languageyou’re speaking

Every other engine makes you pick a language per recording. That breaks the moment someone starts a sentence in one language and finishes it in another, which for bilingual speakers is most sentences.

Seta reads every language it supports in one shared output space, so switching costs it nothing. There is no language setting to get wrong, because there is no language setting.

On the same Kazakh audio, Whisper scores 44.95% word error rate. Seta scores 8.71%.

What it hears todayThe STT reference →
Languages
Kazakh and Russian
Switching
Mid-sentence, and inside a single word
Accuracy
8.71% word error rate, 5× better than Whisper
Speed
40× faster than real time
Model
seta-kk-ru-v2
Next
More languages on the same endpoint

Need a voice that doesn’t exist yet?

Tell us the language, we record the speakers and train the model