Voice AI: when is the "Her" moment?
summary
Claims that full‑duplex fixes everything meet specifics: Moshi full‑duplex demonstrates simultaneous listen/speak and robust overlap handling but delivers little actual utility because it lacks tool integration and paralinguistic exploitation. Cascaded stacks remain practical but suffer from 500ms–4s tool latency that kills conversational timing. TTS cost eats startups’ runway. The underrated technical gap is paralinguistic understanding, not voice quality. A concrete mitigation is Phoneon on-device TTS — <100M-parameter CPU models running locally for privacy and scalable cost.
AI-written summary · about
chapters (13)
- 0:00Start
- 0:14Introduction and mission of Gradium AI
- 1:16Demonstration of voice cloning technology
- 2:42The "Her" movie analogy and current limitations of Voice AI
- 5:42Challenges of cascaded systems (Speech-to-Text, LLM, Text-to-Speech)
- 6:37The difficulty of latency in tool calling
- 9:08Explanation of Speech-to-Speech vs. cascaded architectures
- 9:34The necessity of full-duplex systems and backchanneling
- 11:53Demonstration of the full-duplex Moshi model
- 12:59The importance of paralinguistic understanding
- 14:29Scalability and the high cost of current Voice AI
- 16:38Introducing Phoneon: on-device, local TTS for privacy and cost efficiency
- 18:29Conclusion and path forward for Voice AI
related talks
52mGoing Full Duplex - Building Apps for Speech
28mWhat’s Needed in Tech to bring Voice to the Future
17mBuilding Effective Voice Agents
33mDesigning Voice Agents for Real Conversations
17mReimagining Customer Experiences Integrating Voice into Digital Strategy
13mVoice In, Visuals Out: The Agony and the Ecstasy
51mCan AI Face Your (Potential) Customers?
20mBuilding the next interface - AI personas who feel as natural as interacting with a human