v0.7.2

Expensify/Appv0.7.2May 25, 2026by KoljaB

AI Summary

This release focuses on improving the stability and performance of PocketTTS by reworking the engine to use a persistent worker thread and optimizing memory usage on Windows. It introduces new configuration options for streaming and voice state caching.

Key Highlights

Reworked `PocketTTSEngine` to use a persistent worker thread
Added streaming controls (`streaming`, `max_tokens`, `frames_after_eos`)
Added support for custom PocketTTS model configs and cached voice states
Patched short-text streaming to decode serially, preventing memory retention issues

New Features

Persistent worker thread for PocketTTS synthesis
Streaming controls (max_tokens, frames_after_eos)
Custom PocketTTS model config support
Cached voice state support

Full Release Notes

# RealtimeTTS v0.7.2

### Improvements

* Reworked `PocketTTSEngine` to run synthesis through a persistent worker thread.
* Patched PocketTTS short-text streaming to decode serially, avoiding Windows/Torch CPU memory retention during repeated short generations.
* Added PocketTTS streaming controls: `streaming`, `max_tokens`, and `frames_after_eos`.
* Added support for custom PocketTTS model configs via `model_config`.
* Added support for cached voice states through `PocketTTSVoice(state_path=...)`.
* Added optional voice-state caching through `voice_cache_dir` and `cache_voice_states`.
* Improved PocketTTS audio conversion, queueing, duration tracking, and shutdown cleanup.

### Packaging

* Bumped package version to `0.7.2`.
* Set the PocketTTS dependency to `pocket-tts>=2.1.0`.
* Refreshed explicit dependency versions, including `stream2sentence`, `openai`, `elevenlabs`, `edge-tts`, `requests`, `cartesia`, `camb-sdk`, `omnivoice`, and `typecast-python`.

### Documentation

* Added PocketTTS runtime notes for the persistent worker and serial streaming behavior.
* Cleaned up the README support section.