Notable changes in version 2:

• switch to ABI specific APKs (reduces APK size from 162MB to 76MB for 64-bit ARM)
• fix ONNX resource leaks and refactor usages of ONNX APIs
• handle null TTS voice name to avoid crashes with mismatched languages
• fix phonemizer crashes and text handling on malformed input including edge cases for Markdown links
• fix US abbreviation pronunciation via dictionaries

A full list of changes from the previous release (version 1) is available through the Git commit log between the releases.

GrapheneOS Speech Services provides a built-in text-to-speech implementation for GrapheneOS using a fully open source model for English (US) meaning fully open source training code/data. In the future, it will be expanded to other languages and will also provide speech-to-text. The models built into the app included in GrapheneOS must be fully open source but it can be extended to support additional choices for models distributed through our App Store as additional packages without the same constraint.

Major improvements to performance and the quality of the output will be provided in near future releases. The output is currently slightly distorted by one of the audio processing steps which can be fully removed and replaced by training a new model without a dependency on it. The performance can also be heavily improved by removing or optimizing the audio processing followed by implementing hardware acceleration for the model.

Speech Services should be installed from our App Store which can be installed outside GrapheneOS from GitHub.