Open source · Apache 2.0
Scylla’s Band
Text-to-speech that runs on the phone or computer it’s installed on. No server, no account, no per-character bill.
Ten voices cover American and British English, Spanish, Italian, French, German and Vietnamese, with four delivery controls: energy, tension, valence and assertiveness. Audio streams as it’s generated, so long text starts playing right away.
- 10 voices
- 7 locales
- 24 kHz audio
- 4 delivery controls
- MacBook Pro M5 MaxCore ML, GPU220×
- iPhone 17 Pro MaxCore AI, GPU92×
- MacBook Pro M5 MaxLiteRT, CPU34×
- Galaxy Z Fold 8LiteRT, CPU27×
Approximate. From the model card.
$ git clone https://github.com/lowkeytea/scyllasband
$ cd scyllasband
$ python3 -m scyllasband download
$ python3 -m scyllasband speak --voice scylla \
-o hello.wav "Hello from Scylla's Band."- LiteRT 164 MB
- Linux, Windows, Android and Intel Macs
- Core ML 142 MB
- Apple silicon Macs, iPhone and iPad on iOS 18 and later
- Core AI 155 MB
- iOS, iPadOS and visionOS 27 apps
- ONNX Runtime 262 MB
- Anywhere ONNX Runtime runs
There’s a Python package and CLI, a C++ library with a C API, and sample apps for iOS and Android. It still trips on the odd word, and a few voices get rough on fast, wide pitch jumps. It speaks in its own ten voices only and can’t clone anyone else’s.