Deep Dive
HayaKoe is a TTS engine built by trimming Style-Bert-VITS2 down to Japanese-only and restructuring it into a practical form for CPU inference and server operations.
This section documents what was changed, how it was done, and how much the results improved, with actual measurements.
You can start from any topic that interests you.
Summary at a Glance
Here are the measured improvements HayaKoe achieved over the original SBV2 (see individual pages for details).
| Category | Original SBV2 | HayaKoe | Difference |
|---|---|---|---|
| CPU speed (short text, ~2s) | 1.13 s | 0.68 s | 1.67x faster |
| CPU speed (medium text, ~8s) | 3.35 s | 2.44 s | 1.37x faster |
| CPU speed (long text, ~38s) | 35.33 s | 10.43 s | 3.39x faster |
| CPU memory | 5,122 MB | 2,346 MB | 54% reduction |
| GPU VRAM | 3,712 MB | 1,661 MB | 55% reduction |
| Supported architecture | x86_64 | x86_64 · aarch64 Linux | ARM board support |
Page Structure
Each page follows a why it matters -> implementation -> improvement results flow as a baseline, with flexible structure depending on the topic.
Table of Contents
Big Picture
- Architecture Overview — Full structure of the TTS engine
Making CPU Inference Real-time
- ONNX Optimization / Quantization — Q8 BERT + FP32 Synthesizer, arm64 support
- Sentence Boundary Pause — Duration Predictor — Restoring natural pauses in multi-sentence synthesis
Further GPU Inference Optimization
- BERT GPU Retention & Batch Inference — Eliminating PCIe round-trips and multi-sentence batching
Operational Convenience
- Source Abstraction (HF / S3 / Local) — Unifying speaker sources under URIs
- OpenJTalk Dictionary Bundling — Eliminating first-import delay and network dependency
- arm64 Support — Raspberry Pi 4B measurements
Other
Recommended reading order
If this is your first time, we recommend skimming Architecture Overview first, then selectively reading topics that interest you.