Scylla's Band

Scylla's Band

A text to speech model that also sings from a melody score, with fourteen voices, six languages and controls for intensity, pace and pitch.

Made by spybyscript on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Scylla's Band is a text to speech model that speaks and sings. One model renders both, and a delivery mode picks which. It has fourteen voices and lists English, Spanish, Italian, French, German and Vietnamese. The release described here is v4-20261106.

Singing always needs a score you provide. The model does not make up a song on its own.

Who it suits: It suits developers and creators who want a voice that runs on their own devices, in an app or on a desktop. The runtime covers Linux, Windows, macOS, iOS, visionOS and Android.

Industry | Who uses it | How they use it | Example
App development | Mobile and desktop developers | Add spoken lines to an app | A voice reading a greeting
Games and media | Creators | Give a character a spoken or sung line | A short pop phrase sung from a score
Language learning | App builders | Speak lines in more than one language | A phrase read in Spanish
Music | Songwriters | Hear a lyric set to a melody | A phrase rendered from a score

Skip it if you only need a hosted voice service and never run models yourself.

Where to use it:

• The Scylla's Band runtime, run from Python on a computer
• Linux, Windows and macOS for the command line
• iOS, visionOS and Android through the same runtime
• macOS with the Apple bundles, using the coreml or coreai flavor

Use case | Who | What they get back
Speak a line | Developers | A WAV file in the voice and language you chose
Sing a score | Creators | A WAV file sung in pop or ballad style
Tune the delivery | Anyone | Calmer or more intense, slower or faster, higher or lower speech

How to set it up: You need a few things in place first.

• Python on your computer
• The runtime, installed with pip from a checkout of its code
• Disk space for the voice bundle

1. Install the runtime with pip from a checkout.
2. Run the download command. It installs the bundle and the example scores from the release.
3. On macOS, add the coreml or coreai flavor to fetch the Apple bundles.
4. Optionally run build-native, which makes the commands faster and lighter.

pip install -e .
python -m scyllasband download --flavor litert

You can also get the repository files with Hugging Face's own steps.

pip install -U huggingface_hub
hf download spybyscript/scyllasband

How to use it: You run the speak command with a voice, a language and your text, or the sing command with a voice and a score file. Intensity, pace and pitch each run from minus 2 to 2, and 0 leaves the voice as it is. Values outside that range are clamped.

python -m scyllasband speak --voice ariadne --language en_us --intensity 1 --pace -0.5 \
    -o hello.wav "It is a pleasure to meet you."
python -m scyllasband sing --voice gwen --score scyllasband/models/melodies/phrase_pop_00.json -o ring.wav

Tool | What it does
download | Installs the voice bundle and the example scores
speak | Renders text to a WAV file with the chosen voice and controls
sing | Renders a melody score to a WAV file in pop or ballad style
build-native | Compiles a native version of the runtime for faster, lighter runs

Example: You ask the speak command to say "It is a pleasure to meet you." with the ariadne voice in en_us, at intensity 1 and pace -0.5. It writes hello.wav, a spoken line that is a little more intense and slower than the voice's default.

For newcomers and experts: Start with the download command and the speak command from the setup, then open the WAV file it writes. Change intensity or pace by one step and listen to the difference. When you want singing, try one of the example scores that the download installs. The folder holds twenty of them.

A song is a JSON score with the lyric text, a tempo, a style of pop or ballad, and notes given as label, pitch and beats. Speech and singing share one flow model. Sung output goes through a BigVGAN v2 decoder tuned on the voices, and speech uses Vocos. The flow runs 4 Euler steps. The v4 graph takes an extra delivery_mode_id input, so runtimes built for v3 refuse this bundle.

Install the runtime with pip
Download the voice bundle
Pick a voice and the controls
Speak text or sing a score
A WAV file is written

Good to know:

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: it downloads the bundle and example scores from the repository, then writes the audio file you name
• Not for: free singing without a score, since a melody score is required

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: litert
• Downloads in the last 30 days: 929
• Likes: 17
• License: apache-2.0
• Languages: en, es, it, fr, de, vi
• Tags: litert, tflite, onnx, text-to-speech, singing, singing-voice-synthesis, coreml, pytorch, multilingual, expressive-tts, duration-flow

Links

• Model: https://huggingface.co/spybyscript/scyllasband
• Files: https://huggingface.co/spybyscript/scyllasband/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download spybyscript/scyllasband

Then follow the usage in its model card. Nothing in this kit runs by itself.