Supertonic is a lightning-fast, on-device, multilingual text-to-speech (TTS) system designed for local inference with minimal overhead. Leveraging ONNX Runtime, it operates entirely on the user's device, ensuring privacy and eliminating the need for cloud services or API calls.

Key features include:

• Blazing Fast Synthesis: It offers low-latency, real-time audio generation across various platforms, including desktop, browser, mobile, and edge devices, capable of converting entire webpages to audio in under a second.
• 31-Language Multilingual Support: Supertonic synthesizes speech directly from text in 31 languages. It also provides a language-agnostic mode ($lang="na"$) to process text when the input language is unknown, requiring no separate language adapters.
• Compact Model Size: With an open-weight model of approximately 99 million parameters, it is significantly smaller than 0.7B–2B class open TTS systems, leading to smaller download sizes, faster cold starts, and lower memory footprint.
• Edge-Device Ready: The system runs locally on diverse hardware, from desktops and mobile phones to resource-constrained devices like Raspberry Pi and e-readers, without requiring network dependency or a GPU for fixed-voice inference.
• High-Quality Audio Output: It generates studio-grade 44.1kHz 16-bit WAV audio directly, suitable for production without external upsampling.
• Expression Tags: Supertonic incorporates 10 inline tags (e.g., $<laugh>$, $<breath>$, $<sigh>$) to inject natural human nuances into generated speech without complex prompt engineering or reference audio.
• Multi-Runtime SDKs: Ready-to-use examples and SDKs are provided across various programming environments, including Python, Node.js, Browser (WebGPU), Java, C++, C#, Go, Swift, iOS, Rust, and Flutter.

Technical Details and Core Methodology:
Supertonic utilizes ONNX Runtime for cross-platform inference, with onnxruntime-web enabling client-side inference in browsers. It supports batch inference for improved throughput. The system's core architecture and underlying technologies are based on several research papers:

• SupertonicTTS: Main Architecture: This paper describes the overall system architecture, which includes a speech autoencoder and a flow-matching-based text-to-latent module, emphasizing efficient design choices for on-device deployment.
• Length-Aware Rotary Position Embedding (LARoPE): This technique improves text-speech alignment within cross-attention mechanisms, crucial for accurate and natural speech synthesis.
• Self-Purifying Flow Matching: This method enables robust training of flow matching models even with noisy or unreliable labels, enhancing the quality and stability of the generated speech.
• RobustSpeechFlow: This approach focuses on learning robust text-to-speech trajectories through augmentation-based contrastive flow matching, thereby improving the overall robustness and quality of the generated audio.

Performance Highlights:
Supertonic 3 demonstrates competitive reading accuracy on the Minimax-MLS-test benchmark, achieving Word Error Rate (WER) or Character Error Rate (CER) scores comparable to much larger models like VoxCPM2, OmniVoice, and Qwen3-TTS. For instance, in English, Supertonic 3 achieves a WER of 2.06, competitive with VoxCPM2 (2.11) and OmniVoice (2.02). It runs efficiently on CPU and exhibits a substantially lower memory footprint compared to larger baselines often measured on A100 GPUs. Its compact 99M-parameter model significantly reduces download size and startup time.

Natural Text Handling:
A notable strength of Supertonic is its ability to handle complex, real-world text inputs without requiring pre-processing or phonetic annotations. It excels in:

• Financial Expressions: Correctly pronounces decimal currency ($5.2M as "five point two million"), abbreviated magnitudes (M, K), and currency symbols.
• Phone Numbers: Accurately reads area codes in parentheses, hyphenated numbers, and abbreviated extensions (e.g., "ext. 402").
• Technical Units: Interprets decimal numbers with units (e.g., "2.3h" as "two point three hours") and abbreviated technical notations (e.g., "30kph" as "thirty kilometers per hour").

The project also offers a "Voice Builder" for users to create custom voice profiles for local deployment, generating version-specific JSON files. For managed content creation, zero-shot voice cloning, and API-based integration, Supertonic 3 is integrated into Supertone Play and the Supertone API.

Supertonic is released under the MIT License for its sample code, with the accompanying model under the OpenRAIL-M License.