Reza2kn/gooya-bozorg-v1.5-native

🗣️ Gooya Bozorg 1.5 — Native Runtime Bundle

Fully on-device Persian (Farsi) TTS runtime assets for the Gooya Bozorg 1.5 voice.

🧠 Runtime	ONNX Runtime (T3 + flow-step) · tract (vocoder)
⚖️ Weights	Q4 int4 (MatMulNBits, block 32), ~430 MB
🪣 Bucket	b168 (≤168 speech tokens)
🍎 Platform	macOS / Linux / Windows (CPU + per-OS GPU)
🎯 Latency	~12 s canonical prompt · Apple Silicon CPU

---

📦 Contents

tract-bundle-b168/
├── t3-prefill.onnx / t3-decode.onnx        # speech-token transformer (Q4)
├── s3-flow-prepare-b168.onnx                # flow conditioning (folded)
├── s3-flow-step-b168.onnx                   # flow matching denoiser (Q4)
├── s3-vocoder-source-b168.onnx              # HiFT excitation (Q4)
├── s3-vocoder-spectral-b168.onnx           # spectral head (Q4)
└── *.data                                  # external weights
grapheme_mtl_merged_expanded_v1.json          # grapheme BPE tokenizer

🚀 Use it

# pure Rust asset fetcher (no Python needed)
cargo run --release --manifest-path desktop/Cargo.toml --bin gooya-fetch-assets
cargo build --release --manifest-path webview/Cargo.toml
./webview/target/release/gooya-native-webview

# or build the installable macOS app
./scripts/bundle_app.sh

Full docs: github.com/Reza2kn/gooya-bozorg-native

📜 License

• Model: Reza2kn / Gooya Bozorg 1.5 — CC-BY-NC-4.0
• Runtime bundles: CC-BY-NC-4.0

Non-commercial research/creative use only. 🎨