AI BriefingScylla's Band v4: Open-source TTS model with singing capabilities and 131M parameters
·2026.11.07 07:26
Key point
The 131M parameter model supports 6 languages and 14 voices, achieving 30ms warm start on M5 Max and 0.85% WER on the Open TTS Leaderboard.
Details
- Scylla's Band v4 is an Apache 2.0 licensed open-source TTS model with ~131M parameters (plus a 30M phonemizer) that supports speech and singing.
- The model achieves 0.85% English WER and 4.35 UTMOS on the Open TTS Leaderboard, with warm first audio latency of 30 ms on M5 Max, 70 ms on iPhone 17 Pro Max, and 120 ms on Galaxy Z Fold 8 (CPU only).
- It supports 14 voices across English (US/UK), Spanish, French, German, Italian, and Vietnamese, with per-line controls for intensity, pace, and pitch.
- Weights are available in PyTorch, LiteRT, Core ML, Core AI, ONNX Runtime, and WebGPU formats, alongside a C++ runtime with a C API for text normalization.
- Limitations include the inability to clone voices, occasional errors on specific words, and rough audio on fast pitch jumps during singing.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Craft Anything: Open-source Rust engine for AI-driven software reverse engineering and recreation