v0.8.8

# RealtimeTTS 0.8.8

This release adds optional Qwen CPU decoder overlap: acoustic-code generation can continue while the preceding chunk is decoded, using a separate CPU worker pool. Worker counts, chunk cadence and CPU affinity are configurable.

The tested Linux CPU configuration reduced RTF by 23-26%, with identical PCM in the paired cases. Results depend on hardware and settings. Optional startup priority recovered 18-23 ms of audible-start latency in the measured cases, with a 2-6% generation-time cost.

The full editable emotional demo is restored in tests/faster_qwen_emotions.py. Compact colored output is the default; --verbose enables native diagnostics. Compact output now shuts down cleanly when a child process retains the stderr pipe.

Qwen extras select Numba 0.62.x on Intel Macs to retain binary dependency installation. Intel Mac installations use Python 3.10-3.13. Windows, Linux and Apple Silicon dependency selections are unchanged.

CPU extras pin realtimetts-qwen-native-cpu 0.4.0. Serial/unpinned defaults and the 160 ms startup reserve remain unchanged. Models, precision and sampling are unchanged. GPU native remains at 0.2.0; the experimental GPU startup schedule is not included.

Version 0.8.7 remains a TestPyPI rehearsal version; 0.8.8 includes the installation and shutdown fixes found during that rehearsal.

Install: pip install --upgrade "realtimetts[qwen-cpu]==0.8.8"

For the CPU server: pip install --upgrade "realtimetts[qwen-cpu-server]==0.8.8"