KAME
KAME adapts Kyutai's Moshi speech model to receive guidance from a separate language model while producing real-time spoken responses.
Research release
The speech component is trained to accept guidance
The public workflow includes preprocessing, fine-tuning and checkpoint conversion. Text from an oracle guides upcoming speech during training. This trained component distinguishes KAME from an unchanged speech model connected to another API.KAME fine-tuning workflow and Kyutai initialization
The spoken response and the knowledge source are separate
The inference system runs the speech model alongside a back-end text model. Its public English path also uses speech recognition. Changing the external oracle does not create a new KAME weight release.KAME inference repository
Code licenses do not establish the weight license
Practical guidance. Verify checkpoint terms before redistribution. Also record the oracle and speech-recognition services when measuring latency, cost or data handling.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | SakanaAI/kame |
| Base | Kyutai Moshi; the English training example initializes from moshiko-pytorch-bf16.KAME fine-tuning workflow and Kyutai initialization |
| Weight terms | The reviewed checkpoint card does not state a weight license.kame official model card |