Kokoro local TTS produces garbled speech with WebGPU on AMD Radeon 890M (works correctly with WASM)

[AIWerkstatt]

Summary

On a Linux system with an AMD Radeon 890M (Ryzen AI HX370), Liteforms' local Kokoro TTS produces unintelligible/"mumbled" speech when using the WebGPU backend.

The exact same model produces perfectly clear speech when forced to use the WASM backend.

This appears to be specific to the browser/WebGPU inference path rather than Kokoro itself.

Environment

Hardware
Minisforum AI X1 Pro with AMD Ryzen AI 9 HX370 and AMD Radeon 890M (gfx1150).

Software
Fresh installation (July 15th, 2026) of Ubuntu 24.04 LTS (Wayland, Linux 6.17.0-40-generic).

sudo apt install mesa-utils vulkan-tools mesa-vulkan-drivers
wget https://dl.google.com/linux/direct/google-chrome-stable_current_amd64.deb
sudo apt install ./google-chrome-stable_current_amd64.deb

Run Liteform online in Chrome

google-chrome https://liteforms.lookingglassfactory.com/
--ozone-platform=x11
--enable-unsafe-webgpu
--enable-features=Vulkan,VulkanFromANGLE,DefaultANGLEVulkan
--use-vulkan
Test voice provides garbled speech!!!

See file created from chrome://gpu for more details.

about-gpu-2026-07-16T06-11-40-383Z.txt

Run liteforms-web GitHub

sudo apt install git npm
git clone https://github.com/Looking-Glass/liteforms-web
cd liteforms-web
npm install
npm audit fix
npm run dev
google-chrome http://localhost:3000/
--ozone-platform=x11
--enable-unsafe-webgpu
--enable-features=Vulkan,VulkanFromANGLE,DefaultANGLEVulkan
--use-vulkan
Test voice provides garbled speech!!!

Fix audio bug with Kokoro local TTS

nano workers/kokoro.worker.ts
function getKokoro(
...
const key = ${request.model}:${request.dtype}:wasm;
...
const created = KokoroTTS.from_pretrained(request.model, {
...
device: "wasm",
...
npm run dev
google-chrome http://localhost:3000/
--ozone-platform=x11
--enable-unsafe-webgpu
--enable-features=Vulkan,VulkanFromANGLE,DefaultANGLEVulkan
--use-vulkan
Test voice provides clear speech!!!