Scylla's Band v4 launches multilingual singing TTS
Developer Scylla's Band released v4 under the Apache 2.0 license, publishing PyTorch weights, source code, and ports for LiteRT, Core ML, Core AI, ONNX Runtime, and WebGPU. The approximately 131M-parameter model adds a 30M-parameter phonemizer, 14 voices, seven language options, streaming, and controls for intensity, pace, pitch, and multi-voice dialogue. Its singing format accepts JSON scores in pop or ballad styles but currently supports English only; the developer reports that some voices sound rough on fast, wide pitch jumps.
Scylla's Band v4 combines compact multilingual speech, singing, released model weights, and cross-platform runtimes in an Apache 2.0 project, although its quality claims and benchmarks remain developer-reported.