Subnet Session with Koyuki from Vocence: Subnet 78

This episode, Mark discloses that DSV is already invested in today’s subnet, but they’ll still ask the awkward questions. They bring on Koyuki (“special k”) from San Francisco, who shares her background in AI (web2 + web3), how she joined the Bittensor Foundation/OTF as Head of AI, and then dives into her slides on Subnet 78, Vocence.

Koyuki pitches Vosens as a decentralized “voice intelligence layer” on Bittensor, targeting the rapidly growing voice AI market and competing with incumbents like ElevenLabs by being more open, cheaper, and driven by Bittensor incentives. She shows that Vocence already has a live studio product (TTS/STT, voice cloning/design, text-to-music, API) and outlines how miners submit models that validators score across nine dimensions (script accuracy and naturalness weighted highest), with winning models becoming the new baseline for inference. On revenue, she describes a credit-based SaaS model (consumer + API, with enterprise as the big upside), plans for buybacks into a treasury, and an emissions burn condition if no model clears a defined improvement threshold. The discussion then focuses on the “Turing test” problem for voice agents—latency, filler words, interruptions, and overlapping speech—and Koyuki claims a new “style trajectory TTS” approach will make agents sound truly human soon. Siam offers a $5,000 wager that Vocence can produce a voice agent he can’t detect as AI by the end of the month, and Koyuki accepts, with some talk about testing via a phone-call scenario and adversarial off-script questions. They wrap by noting the prior Vocence slot issues/deregistration risk and arguing this time is different due to stronger leadership, a live product, faster shipping, and early traction.