v0.8.10

RealtimeTTS 0.8.10 improves Qwen streaming startup and removes CPU quiet-onset retries.

- CPU synthesis now continues through initial quiet audio and trims it in the original stream, without cancelling and generating the fragment again.
- Voice Studio defaults to **0 ms additional browser buffering**, configurable in the playback options. It prepares and reuses the audio context before speech requests.
- Server logging runs outside the request event loop. The CPU extra now uses native 0.4.2, which also routes prompt diagnostics through the native logging callback.
- CPU and NVIDIA GPU server extras include the optional `--demo-voice` cloning example.
- The new opt-in `--cpu-core-split` discovers Windows physical cores and places generation and decoder workers separately, without raising process priority.

Windows CPU (example for 12 physical cores):
```bat
uv venv --python 3.11 --managed-python .venv-cpu
uv pip install --python .venv-cpu\Scripts\python.exe "realtimetts[qwen-cpu-server]==0.8.10"
.venv-cpu\Scripts\realtimetts-qwen-server.exe --device cpu --cpu-threads 6 --cpu-codec-threads 6 --cpu-core-split --cpu-stream-frames 1 --cpu-fused-attention --startup-buffer-ms 80 --onset-silence-profile qwen3_tts_12hz_0_6b_base_q8_v1 --demo-voice
```

This uses six generation workers and six decoder workers with an 80 ms server reserve. Adjust both counts to the available physical cores. On hybrid CPUs, core splitting uses the highest performance core class; too few cores produce a clear error. The short `--device cpu --cpu-fused-attention` command alone does not enable decoder overlap. Core splitting is Windows-only; omit that flag on Linux/macOS.

Windows NVIDIA GPU:
```bat
uv venv --python 3.11 --managed-python .venv-gpu
uv pip install --python .venv-gpu\Scripts\python.exe "realtimetts[qwen-server]==0.8.10"
.venv-gpu\Scripts\realtimetts-qwen-server.exe --device gpu --clone-mode speaker_only --demo-voice
```

Open http://127.0.0.1:8080/studio and select **demo-neutral**. The GPU command uses the existing 160 ms server reserve. Server reserve and browser buffering are separate, configurable controls.

See the [installation and tuning guide](https://github.com/KoljaB/RealtimeTTS/blob/v0.8.10/docs/qwen-studio.md) for manual CPU settings and voice cloning. Performance depends on hardware, text, voice, and concurrent load.