- Who it helps
- Researchers studying turn-taking and audio models
- Availability
- November 10 research description and Apache-2.0 code are public; model-card access was unavailable during this review.
What the experiment tests
Two audio streams represent the participants. The 413-million-parameter model is intended to explore the rhythm of conversation before reliable language. It is unsuitable as evidence of a useful voice assistant.
Available evidence
The creator reports faster-than-real-time performance on a GPU and an M3 Pro. Those measurements were not reproduced, and the linked audio sample was not played.
What you can do
- Inspect the creator’s semantic limitations before comparing this with a voice assistant. Open reference ↗
Where the evidence stops
Descriptions and code documentation checked November 11, 2026. No audio generation, listening evaluation or latency test performed.
Sources read. Public source text checked November 11, 2026 UTC. No third-party project or complete workflow executed.
Sources & further reading
Web · official
Caramel-1 speech experiment ↗
November 10 description separates conversational timing from meaningful language.
naklecha · page checked 2026-11-11 · Published 2026-11-10
GitHub · implementation
Caramel implementation ↗
README and Apache-2.0 code inspected; audio was not played and weights were not loaded.
naklecha · page checked 2026-11-11
4 stars · observed 2026-11-11 06:28 UTC
Coverage note: Descriptions and code documentation checked November 11, 2026. No audio generation, listening evaluation or latency test performed.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Caramel-1 is a small speech-only experiment that learns conversational timing, pauses and overlapping turns. The creator says most of its words are not meaningful.