v0.8.9

# RealtimeTTS 0.8.9

Qwen CPU and GPU servers are easier to install and use, with corrected streaming playback in Voice Studio.

- Fix standalone synthesis failures for language codes such as `de` and `en`, and preserve inherited codec sampling.
- Avoid artificial rebuffering at contiguous audio boundaries. Studio uses an editable 80 ms browser buffer for CPU and GPU and reports actual playback gaps.
- Open Studio in English. `--demo-voice` downloads, verifies, registers and warms a public neutral cloning example.
- Provide the installed `--preset windows-3900x` profile: six generation workers, six decoder workers, one-frame chunks and an 80 ms server reserve.
- Pin CPU native 0.4.1. The separate `--cpu-fused-attention` flag enables its optional F32 attention path. GPU native remains 0.2.0.

The optional CPU path keeps model weights, quantization and sampling settings. Floating-point evaluation order changes, so seeded audio need not be identical. On the tested Ryzen 3900X, a paired English-sentence comparison reduced mean RTF by about 6.4%. The final Windows CPU and GPU browser listening checks recorded zero underruns. Earlier CPU runs under changing load still had gaps; an 80 ms reserve does not guarantee uninterrupted playback for every text or competing workload.

Install `realtimetts[qwen-cpu-server]==0.8.9` for CPU or `realtimetts[qwen-server]==0.8.9` for NVIDIA GPU. Follow the [three-command uv quick start](https://github.com/KoljaB/RealtimeTTS/blob/v0.8.9/docs/qwen-studio.md#quick-start-with-a-cloning-example).

No fresh deployment to the production Linux machine is claimed.