Kotoba CEO: Audio Tokenization Is the Core Research Problem in Real-Time Speech AI
lmthang · x · 2026-11-22
A repost of notes from a guest lecture by Noriyuki Kojima, co-founder and CEO of Kotoba (which recently joined ElevenLabs). He shared his path from building Fugaku-LLM, Japan's own LLM, to founding Kotoba, plus technical insights into frontier real-time speech AI.
Key points:
- Audio tokenization is an important research problem, with a trade-off between compressing the audio representation and preserving acoustic information that affects training and inference efficiency.
- Use case dictates the compression strategy: call centers care most about semantic content and can compress aggressively, while music generation must preserve acoustic detail.
- The lecture was hosted with GStar bootcamp 2025 alumni.
More from Multimodal
- AI creator shows off same Viking warrior character in 2D vs 3D styles — azed_ai · 2026-11-22
- Creator releases 120+ free customizable video components for AI-generated GTA 6 videos — eptwts · 2026-11-22
- Magnific One demo: 10 illustration styles with full prompts in thread — charis_ai · 2026-11-22
- An Actor Licensed His Face to a Microdrama Startup. Is This Hollywood's AI Future? — TechCrunch AI · 2026-11-22
- AI video nails Bruce Lee martial arts with fluid fight choreography — mark_k · 2026-11-21
- Founder discovers Plotzy, a TikTok AI short-drama creator monetizing via subscriptions — NathanSRobinson · 2026-11-21