Main changes:
- The NLLB translation model has been replaced with the option to choose between: Mozilla Bergamot models, Madlad 400 3B, and HY-MT 1.5 1.8B. All these models have better translation quality than NLLB 54B, similar to 2024's Google Translate, with HY-MT 1.5 1.8B having much better scores than Google Translate, similar to DeepSeek v3 and Gemini 3.0.
- MLKit has been replaced with open-source alternatives (FastText and Lingua) for language detection, making RTranslator 100% open source.
- Various techniques have been added to improve translation quality, including: better beam search, multilingual dictionaries (combination of Wiktionary, FreeDict and Apertium), Tatoeba integration, and more.
- Improved TTS for WalkieTalkie and Conversation mode: now each message has a button to start and stop the TTS for that message. (Thanks to @NikolaCavendish, PR #183).
- Implemented an audio queue for WalkieTalkie mode: now the mic is deactivated only during the TTS reproduction. So with TTS deactivated you can talk and translate continuously and even when it is activated, the TTS will start for all the new messages only after no speech is detected.
Other changes:
- Improved speech detection for the automatic mic in WalkieTalkie and Conversation mode with a mix of volume level (the old way) and a small neural network that detects voice activity (Silero VAD via Android VAD).
- Total rebuild of the download system, now based on PRDownloader, increased stability, added support for pause, improved notification.
- Improved Bluetooth stability for Conversation mode.
- Added a description dialog before any permission request.
- Updated Android APIs to 36.
- Better code organization based on git submodules for external libraries.
Warning: This is a beta release, so it can have a lot of bugs! If you want a stable experience, you should use the latest stable version and wait for the final stable release of the 3.0 version (a few months from now).
This release is funded through the NGI Mobifree Fund, a fund established by NLnet. A special thanks goes to them!
If you also want to support this project yourself, you can buy me a coffee; it means a lot ❤️