Starstruk Logo MarkStarstruk

v0.7.3

agentscope-ai/QwenPawv0.7.3May 25, 2026by KoljaB

AI Summary

Adds GPU support for the PocketTTS engine and optimizes the library's import behavior by delaying tokenizer initialization until actual use.

Key Highlights

  • New `PocketTTSGpuEngine` and `PocketTTSGpuVoice` classes for CUDA devices
  • Delayed initialization of `stream2sentence` tokenizer to reduce import overhead
  • Added explicit `device` support to `PocketTTSEngine`
  • Added `pockettts-gpu` and `pocket-gpu` extras for GPU dependencies

New Features

  • GPU engine classes and built-in voice support
  • Lazy root exports and engine exports
  • New installation extras for GPU dependencies

Full Release Notes

# RealtimeTTS v0.7.3

### Features

* PocketTTS GPU
  * Added `PocketTTSGpuEngine` and `PocketTTSGpuVoice` for the CUDA-capable `pocket-tts-gpu` fork.
  * Added built-in voice support, prompt WAV voice cloning, cached prompt state loading/saving, and optional silence trimming for the GPU engine.
  * Added lazy root exports and `RealtimeTTS.engines` exports for the new GPU engine.

### Improvements

* PocketTTS
  * Added explicit `device` support to `PocketTTSEngine` so CPU users can stay on CPU while advanced users can move the official PocketTTS model to another Torch device.
  * Kept CPU and GPU PocketTTS paths separate so normal `realtimetts[pockettts]` installs remain predictable.
* Import Behavior
  * Delayed `stream2sentence` tokenizer initialization until sentence generation is actually used, keeping `import RealtimeTTS` quieter and lighter.
* Packaging
  * Bumped the package version to `0.7.3`.
  * Added `pockettts-gpu` and `pocket-gpu` extras for the shared GPU-engine dependencies.

### Fixes

* Avoided import-time tokenizer setup from `TextToAudioStream`, which could surprise users before they created or played a stream.
* Improved the missing-dependency message for `PocketTTSGpuEngine` with the pinned GPU fork install command.

### Documentation

* Expanded PocketTTS documentation with separate CPU and CUDA fork installation paths.
* Documented CUDA PyTorch installation and the pinned PocketTTS GPU fork revision used for this release.
* Updated installation and engine-selection docs to list `PocketTTSGpuEngine` and the new GPU extras.
* Updated the README engine matrix to show the optional PocketTTS GPU path.

### Tests and Examples

* Added unit coverage for `PocketTTSGpuEngine` streaming output and `PocketTTSGpuVoice` representation.
* Expanded PocketTTS tests around device forwarding.
* Verified the PyPI wheel and source distribution with `twine check`.
* Smoke-tested the built wheel by installing it and importing `PocketTTSGpuEngine` / `PocketTTSGpuVoice`.

### Notes

* PyPI release: `pip install realtimetts==0.7.3`
* CPU PocketTTS install: `pip install "realtimetts[pockettts]"`
* GPU PocketTTS install starts with `pip install "realtimetts[pockettts-gpu]"`, then requires a CUDA-enabled PyTorch build and the pinned `pocket-tts-gpu` fork documented in `docs/engines/pockettts.md`.
* No intentional breaking changes.