RealtimeTTS 0.8.9

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.

No fresh deployment to the production Linux machine is claimed.