Kugel 3.5 - Text-to-speech that beats any other Model out there

Kugel 3.5 beats four leading TTS models on 8 of 9 tests. It gets short words wrong at half the rate of the next-best model, reads long numbers exactly right 97.1% of the time and supports IPA. 5.25 cents a minute. Not on Kugel? Three lines to production.

Add a comment

Replies

Best
Hey everyone, we're back! This time it's Kugel 3.5, our new frontier TTS model. Nobody wants to talk to a bot that sounds like one. Often the giveaway is small: a phone number read back wrong, or a "yes" that comes out mangled. So we put Kugel 3.5 and four of the leading TTS models through the same 9 tests. Kugel 3.5 came first on 8: it gets short words wrong at half the rate of the next-best model (3% vs 6.1%) and reads long numbers exactly right 97.1% of the time, the highest of all five. It also costs 25% less than our previous model, 5.25 cents a minute. So, if your agents aren't running on Kugel yet, it's three lines to production. Try it on your hardest inputs and tell us where it breaks.

Phone numbers are a good test case for speech tools: one wrong digit changes the whole message. Good to see that getting attention alongside how natural the voice sounds, Kajo.

 thanks Oleksii, one of the many reasons we built this