Summary
According to a post on X by @cartesia, H-Nets, from Albert Gu and the Cartesia team, learn directly from raw bytes and form their own units instead of relying on a fixed tokenizer. The post says a 3B H-Net matched an open-source model while using roughly 10x less training data. Early results on DNA and images are mentioned, and Cartesia sees the approach as a path for speech-to-speech models that learn straight from raw audio. The snippet is truncated, so the full research details are not available here.
AI-assisted summary based on the listed source.
What happened
H-Nets, from Albert Gu and the Cartesia team, learn directly from raw bytes and form their own units instead of relying on a fixed tokenizer. Cartesia says a 3B H-Net matched an open-source model with roughly 10x less training data, with early results on DNA and images, and sees it as a path for speech-to-speech...
Why it matters
Teams building speech and voice models could watch whether tokenizer-free designs that learn from raw audio hold up beyond the early results described. If the claimed data efficiency holds in wider testing, it could reduce training data needs for some models. These claims come from Cartesia's own post, so they should be checked against the full research.
Signal Intelligence
Signal Strength 83%
Technical label WATCH
Public Interest 43
Category OPEN SOURCE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 73
Practical Impact Score 8
Novelty Interest Score 70
Consequence Score 10
Curiosity Score 16
Shareability Score 58
Why this is here
VQV surfaced this signal because it is recent, relevant to AI Voice, connected to Scobleizer News Pulse (X).
