Koyuki Nakamori: Perturb SN26, Vocence SN78, Voice AI & SOTA, Adversarial Robustness | Ep. 93

In this episode, Koyuki Nakamori, CEO of Perturb (Subnet 26) and co-builder of Vocence (Subnet 78), details the development of two complementary decentralized AI networks on Bittensor — an adversarial robustness testing platform and a decentralized voice intelligence layer. She discusses her career trajectory from head of ML at Headspace to the Avalanche ecosystem to dual-subnet founder, the evolution of the Bittensor ecosystem under dynamic TAO, and the strategic relationship between Perturb's perturbation-based data synthesis and Vocence's voice model training pipeline. Koyuki also details Perturb's adversarial image perturbation mechanism for solving the long-tail problem in AI classification, Vocence's nine-dimensional voice quality evaluation system and Qwen3.6-27B miner baseline, the winner-take-all competition structure, and her advice for subnet teams on closing the gap between research and consumer product. *In this episode, you'll learn* - Koyuki's career arc from production ML engineering at Headspace and the Avalanche ecosystem to founding two Bittensor subnets simultaneously across adversarial robustness and voice AI - The mechanics of Perturb's adversarial robustness network, where miners compete to produce maximally deceptive image perturbations within distortion bounds to expose long-tail vulnerabilities in AI classifiers - How Perturb functions as a data synthesis engine for niche AI domains — autonomous vehicles, medical imaging, and precision applications — by adding controlled noise to synthesize training data that real-world acquisition cannot provide - Perturb's multi-modal roadmap, progressing from image perturbations to audio, video, and robotics applications - Vocence's miner incentive model, where miners train voice AI models to surpass Qwen3.6-27B across nine evaluation dimensions — tone, accuracy, naturalness, and accent among them — in a winner-take-all competition for inference serving rights on vocence.ai/studio - The nine standardized

Read the article at Ventura Labs (YouTube) ↗

Taosis indexes announcements, articles, posts and releases about Bittensor subnets from their own channels and keeps a record of each. The text above is the source’s; the figures are read from the chain by Taosis.