Block's speech-only voice AI flatters to deceive: demos contain factual errors
rdesh26 · x · 2026-11-06
Block's Dessa team claimed its new voice AI, trained entirely on speech without text, developed surprisingly strong language understanding. rdesh26 pushes back: checking a movie-dialog demo transcript reveals factual errors — e.g., looking forward to Nolan's "third" Batman film (released 2012), and misattributing Superman IV's direction to Bryan Singer instead of Sidney J. Furie.
- A dialog model trained on text+speech, or with web search, wouldn't have these gaps
- Fixing via human-recorded conversations would cost thousands more dollars; TTS defeats the "natural dialog" purpose
- "Somehow ours ended up better" — better than what, on which metrics?
Context: GSLM and Hertz-dev tried speech-only before; the claimed gap is language understanding.
More from Models
- 12B open Nemotron trained on Surge RL data beats GPT-5.6 on SWE-Bench Pro V2 — echen · 2026-11-06
- Boson AI launches Higgs Realtime Duplex: 45B MoE full-duplex omni model that listens, talks and works simultaneously — smolix · 2026-11-06
- Boson AI's 45B MoE duplex model tops Big Bench Audio at 98.9, beating TML and Google Live — smolix · 2026-11-06
- Artificial Analysis posts Intelligence Index v5 results; Nanbeige4.3-33B scores 20 — ArtificialAnlys · 2026-11-06
- Nanbeige4.3-33B hallucinates less than Qwen3.8 27B but answers fewer correctly — ArtificialAnlys · 2026-11-06
- BOSS Zhipin's Nanbeige4.3-33B open model matches Qwen3.8 27B on AA Index — ArtificialAnlys · 2026-11-06