It's good to see more open models approaching on-device inference. We need more stops on the fast<>good quality spectrum than just piper and VITS.
My first impressions of it:
* The cloning was decent at imitating voices, but the prosody is quite bad
* There's noticeable crackling in the GGUF models and the quality drop from base model to Q8 was significant
* Q4 models are apparently bugged on platforms outside of Linux
* The speed is nowhere near realtime even using all the latency reductions (Q4 backbone, pre-encoding, ONNX codec decoder), it was still lucky to hit a real-time factor of 4x
> Optimised for on-device deployment - provided in GGML format, ready to run on phones, laptops, or even Raspberry Pis
All of this testing was on a beefy 24 core AMD with 64GiB of RAM. There's no way this model would even come close to realtime on any Pi I know.
My first impressions of it:
* The cloning was decent at imitating voices, but the prosody is quite bad
* There's noticeable crackling in the GGUF models and the quality drop from base model to Q8 was significant
* Q4 models are apparently bugged on platforms outside of Linux
* The speed is nowhere near realtime even using all the latency reductions (Q4 backbone, pre-encoding, ONNX codec decoder), it was still lucky to hit a real-time factor of 4x
> Optimised for on-device deployment - provided in GGML format, ready to run on phones, laptops, or even Raspberry Pis
All of this testing was on a beefy 24 core AMD with 64GiB of RAM. There's no way this model would even come close to realtime on any Pi I know.