⚡ Suzune Flash (82M Flagship Multilingual Edge TTS)

Suzune Flash is an 82-million parameter studio-grade multilingual Text-to-Speech (TTS) foundation model engineered by Sorika Labs.

Designed for real-time edge intelligence, Suzune Flash produces broadcast-quality speech with sub-25ms latency on CPUs, Apple Silicon (M1/M2/M3), and GPUs, featuring native English and Hindi studio voices.

🏛️ Base Model & Lineage

Lineage & Attribution: Suzune S1 Flash is developed and released by Sorika Labs. The model architecture is adapted and fine-tuned from the open-source hexgrad/Kokoro-82M and StyleTTS2 foundation (released under Apache 2.0).

Sorika Labs Enhancements & Customizations:

1. Signature Studio Voices: Specially trained and calibrated flagship voices (ira, kian, kagami, phi).
2. Dedicated Hindi Studio Voices: Native Indian speech synthesis with studio voices (ananya, aarav, priya, rohan).
3. Built-in Pure Python Devanagari G2P: Zero-dependency grapheme-to-phoneme engine for pure Hindi Devanagari text (नमस्ते! यह सोज़ूने है...) — works universally across Windows, macOS, Linux, Kaggle, and Colab with zero external C-libraries like espeak-ng.
4. Native ONNX Quantized Engine: Up to 5x faster CPU inference via companion repository sorika-labs/suzune-s1-flash-gguf.
5. Unified SDK: High-level suzune-flash Python runtime with streaming, multi-voice interpolation, and CLI.

🚀 Quick Install

pip install suzune-flash

🎙️ Python Quickstart

1. English Studio Speech

import suzune_flash

# Load Suzune Flash (default signature voice: 'ira')
tts = suzune_flash.load(voice="ira", device="cpu")

audio, sr = tts.speak(
    "Hello! This is Suzune Flash by Sorika Labs, delivering breathtaking studio quality speech.",
    output="english_studio.wav"
)

2. Native Hindi Studio Speech

import suzune_flash

# Load native Hindi female voice 'ananya' or male voice 'aarav'
tts = suzune_flash.load(voice="ananya", device="cpu")

hindi_text = "नमस्ते! यह सोज़ूने फ़्लैश है, जिसे सोरिका लैब्स द्वारा विकसित किया गया है।"
audio, sr = tts.speak(hindi_text, output="hindi_studio.wav")

3. Fast Quantized ONNX Mode (CPU)

import suzune_flash

# Up to 5x faster CPU inference via Q8 ONNX
tts = suzune_flash.load(engine="onnx", quant="q8", voice="ira")
audio, sr = tts.speak("Running fast quantized ONNX speech on CPU.", output="fast.wav")

🎭 Available Voices

💻 Terminal Command Line (CLI)

suzune-flash "Hello from Sorika Labs on-device AI!" --voice ira --output output.wav

📜 License & Citation

Licensed under the Apache 2.0 License.

@misc{sorika2026suzuneflash,
  author = {Sorika Labs},
  title = {Suzune S1 Flash: Flagship Multilingual On-Device Neural Speech Foundation Model},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {https://huggingface.co/sorika-labs/suzune-s1-flash}
}