Moshi team's fun checkpoint: 2k hours of 90s phone calls made every call feel like 1999
rdesh26 · x · 2026-10-24
Moshi co-creator neilzegh recalled that two years ago they trained a fun early checkpoint directly on Fisher (late-90s/early-2000s phone calls) with just 2k hours of data — every call connected you to a random stranger from that era. He notes hundreds of thousands of hours of full-duplex conversational data now exist, so there's an obvious scaling law for full-duplex voice. Context: Jack launched Dessa, billed as the most authentic full-duplex speech model, with emotional, natural human-like cadence.
More from Multimodal
- Kling AI demo shows lifelike female singer performing in a KTV lounge — Sure_Farmer5321 · 2026-10-24
- Creator hooks ChatGPT to Runway via MCP to drive Seedance 2.5 cinematic clips — aziz4ai · 2026-10-24
- Grok Imagine releases Odyssey Episode 1, a ~7-minute AI film with Wonder Studios — lmoroney · 2026-10-24
- ElevenLabs launches Chimera, pairing Eleven v4 voices with Seedance 2.5 to fix AI video's audio tell — umesh_ai · 2026-10-24
- Nano Banana 2.1 renders a cinematic Mars spaceport at nightfall — TinfoilTricorn · 2026-10-24
- Higgsfield Genjutsu character swap tested: MiniMax H3 vs FastH3 Omni on RTX 5090 — CeFurkan · 2026-10-24