⚡ Suzune S1 Flash GGUF / ONNX (Quantized Edge TTS)

This repository provides official quantized, low-latency ONNX and GGUF checkpoints of Suzune S1 Flash engineered by Sorika Labs.

Optimized specifically for real-time edge devices, CPU-only servers, and low-memory environments, achieving up to 5x faster CPU inference with minimal RAM overhead.

---

🏛️ Base Model & Lineage

Lineage & Attribution: Quantized from sorika-labs/suzune-s1-flash, which is built upon the StyleTTS2 and hexgrad/Kokoro-82M open-source foundation (Apache 2.0).

Quantization pipelines (Q8, Q4, FP16), Hindi voice calibration, and the standalone SuzuneFlashONNX runtime were engineered and released by Sorika Labs.

---

📦 Model Variants & Checkpoints

---

🚀 Quickstart via suzune_flash SDK

No third-party wrappers required. Suzune Flash includes native ONNX engine support:

pip install suzune-flash onnxruntime

Python Inference (Q8 Fast Mode)

import suzune_flash

# Load Q8 quantized ONNX engine (English: 'ira', Hindi: 'ananya')
tts = suzune_flash.load(engine="onnx", quant="q8", voice="ira")

audio, sr = tts.speak(
    "Hello! Suzune Flash ONNX runs with ultra-low latency on pure CPU.",
    output="fast_speech.wav"
)

Native Hindi with ONNX

import suzune_flash

tts = suzune_flash.load(engine="onnx", quant="q8", voice="ananya")

hindi_text = "नमस्ते! सोज़ूने फ़्लैश ओएनएनएक्स वर्ज़न सीपीयू पर बहुत तेज़ी से आवाज़ उत्पन्न करता है।"
audio, sr = tts.speak(hindi_text, output="hindi_fast.wav")

---

📜 License & Citation

Licensed under the Apache 2.0 License.

@misc{sorika2026suzuneflashgguf,
  author = {Sorika Labs},
  title = {Suzune S1 Flash GGUF/ONNX: Quantized Real-Time Edge Speech Foundation Models},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{https://huggingface.co/sorika-labs/suzune-s1-flash-gguf}}
}