lowkeytea/scyllasband

Scylla's Band is a runtime package for playing back speech from a text to speech model. It takes text you give it and turns it into an audio file, using a model that was already trained somewhere else. This repository only contains the code needed to run and use that model, not the code that trained it in the first place. The default way it runs is through ONNX, a common format for running trained models efficiently on a desktop or server. There is also an experimental alternative path called LiteRT meant for native and mobile testing, but you have to choose it directly since it is not the default option. The output is 24 kHz audio, and the tool understands four languages: US English, British English, Spanish, and Italian. It ships with ten built in voices with names like Ariadne, Felix, and Scylla, and lets you shape how a line sounds using six independent style controls: calm, joy, anger, sadness, sarcasm, and questioning. These are not fixed categories you pick one of, they are sliders you can mix together, so a line can sound calm and questioning at the same time. A separate setting called CFG scale controls how strongly the requested style pulls the delivery away from a neutral baseline, with the default acting as normal application and higher values pushing further, though very high values can start to sound unstable or distorted. You can also feed it longer scripts, and the tool automatically breaks the text into manageable chunks while keeping track of natural pauses and context between them. A group speak mode lets you write dialogue where each line is tagged with which voice, language, and style mix should read it, which is meant for putting together short scenes with multiple speakers. To try it, you install it with Python's package manager, download the model files with a built in command, then use a command line tool to list available voices or generate speech from text. It is written in Python and is still early stage, with six stars on GitHub so far.