Scylla's Band v4: open-source 131M-param TTS that sings, runs on phones with 0.85% English WER
clockentyne · reddit · 2026-11-07
A solo, year-long project, Scylla's Band v4 is a small open-source TTS model that runs on phones — and can sing from JSON scores (English-only singing for now).
Specs
- 131M params (plus a 30M phonemizer), 14 voices, English (US/UK), Spanish, French, German, Italian and Vietnamese; every voice speaks every language.
- Per-line intensity/pace/pitch controls, multi-voice dialogue with inline tags, streaming; singing via JSON scores with per-word notes and beats, pop or ballad.
Performance
- Warm first audio: 30ms on M5 Max, 70ms on iPhone 17 Pro Max, 120ms on Galaxy Z Fold 8 (CPU only); over 30x real time on desktop CPU.
- English WER 0.85%, UTMOS 4.35 on Open TTS Leaderboard eval (submission pending).
- Apache 2.0, with PyTorch weights plus ports for LiteRT, Core ML, ONNX Runtime and WebGPU, a C++ runtime handling text normalization, and sample apps for Python/iOS/Android.
Limitations: no voice cloning, occasional mispronunciations, some voices get rough on fast wide pitch jumps while singing.
More from Multimodal
- Your LoRA might be fine — it may just be the wrong base model, says SDXL-era trainer — Inevitable_Pen9043 · 2026-11-07
- Game dev uses Scenario's MCP to auto-generate ruins, alien portals and landmarks — AIandDesign · 2026-11-07
- Arena weekly: Kimi K3.1 Pro tops open models, Gemini leads video, GPT Image 2x faster — arena · 2026-11-07
- Sand3D Explorer lets you browse 106K+ 3D objects with parts, mechanisms and code — victormustar · 2026-11-07
- 17 video effects Claude can create from one take, one per second — aakashgupta · 2026-11-07
- "This is just a dream": cinematic AI-generated short film wows Reddit — Sweaty-Ad4488 · 2026-11-07