Wyoming Kokoro Torch

Wyoming protocol server for the original Kokoro Torch TTS implementation.

Contrary to other Wyoming implementation, wyoming-kokoro, this is one uses Torch instead of ONNX. As of the time of writing, our implementation also supports streaming mode, while the ONNX one doesn't. Streaming is important for LLM-based assistant, so that it can start speaking before the LLM is finished generating.

Local Install

Clone the repository and set up Python virtual environment:

git clone https://github.com/debackerl/wyoming-kokoro-torch.cpp.git
cd wyoming-kokoro-torch
script/setup

Download the base model:

mkdir /data
wget -O /data/kokoro-v1_0.pth https://huggingface.co/hexgrad/Kokoro-82M/resolve/main/kokoro-v1_0.pth
wget -O /data/config.json https://huggingface.co/hexgrad/Kokoro-82M/resolve/main/config.json

Run a server anyone can connect to:

script/run --voice af_heart --streaming --uri 'tcp://0.0.0.0:10300' --data-dir /data --download-dir /data

See available voices.

Remarks

If you run this in a VM, you may see Could not initialize NNPACK! Reason: Unsupported hardware. in the logs. This seems to happen on heterogeneous CPU architectures, like my AMD Ryzen AI 7 HX 370. Solutions seems to be either running bare-metal or pass the L3 cache information to the VM and tweak CPU affinity.