Supertone releases Supertonic 3: 99M-parameter on-device TTS model supporting 31 languages
Supertone has released Supertonic 3, a 99M-parameter text-to-speech model that runs entirely on-device using ONNX Runtime. The model expands language support from 5 to 31 languages compared to Supertonic 2, requires no GPU, and claims competitive accuracy against models 7-20x larger.
Supertone releases Supertonic 3: 99M-parameter on-device TTS model supporting 31 languages
Supertone has released Supertonic 3, a 99M-parameter text-to-speech model that runs entirely on-device using ONNX Runtime. The model expands language support from 5 to 31 languages compared to Supertonic 2 and requires no GPU for inference.
Technical Specifications
- Parameters: 99 million across ONNX assets
- Languages: 31 (expanded from 5 in Supertonic 2)
- Inference: CPU-only via ONNX Runtime, no cloud calls required
- Model type: Text-to-speech
- License: OpenRAIL-M for model weights, MIT for sample code
Performance Claims
According to Supertone, Supertonic 3 achieves competitive word error rates (WER) and character error rates (CER) against larger open-source TTS models like VoxCPM2, which range from 0.7B to 2B parameters. The company provides benchmark comparisons showing the model runs faster on CPU than larger baselines measured on A100 GPU.
Supertonic 3 claims improvements over version 2 in three areas: reduced repeat and skip failures during reading, higher speaker similarity across shared languages, and the 6x expansion in language coverage.
New Features
- Expression tags: Supports
<laugh>,<breath>, and<sigh>tags for expressive synthesis - Improved stability: Fewer reading errors on both short and long text inputs
- 31 languages: English, Korean, Japanese, Arabic, Bulgarian, Czech, Danish, German, Greek, Spanish, Estonian, Finnish, French, Hindi, Croatian, Hungarian, Indonesian, Italian, Lithuanian, Latvian, Dutch, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Swedish, Turkish, Ukrainian, Vietnamese
Deployment
The model ships as ONNX assets and runs through a Python SDK. Users can install via pip install supertonic and generate speech locally. The SDK auto-downloads model assets from Hugging Face on first run.
from supertonic import TTS
tts = TTS(auto_download=True)
style = tts.get_voice_style(voice_name="M1")
wav, duration = tts.synthesize(text, voice_style=style, lang="en")
What This Means
Supertonic 3 targets the growing demand for privacy-preserving, on-device AI inference. At 99M parameters, the model is 7-20x smaller than comparable open TTS systems, making it practical for browser and edge deployment where GPU access is limited or unavailable. The CPU-only requirement and sub-100MB footprint address real constraints in mobile and embedded applications.
The 31-language support positions Supertonic 3 as a lightweight alternative to larger multilingual TTS systems. However, without independent benchmarks, it remains unclear how the model's accuracy-size tradeoff compares to cloud-based alternatives or other on-device TTS solutions across different hardware profiles and use cases.
Related Articles
Underdog releases Woof, a 31M-parameter open text-to-speech model that runs in a browser
Underdog's new text-to-speech model Woof has 31 million parameters and weighs 31.7 MB. The company says it runs on most devices without an internet connection, and the weights are open to download. Underdog demonstrated it with a browser game posted on November 5, 2026.
Preferred Networks opens PLaMo 3 Translate weights in four sizes, 0.8B to 31B
Preferred Networks released PLaMo 3 Translate, a translation-specialized LLM, as open weights on November 5 in four sizes: 0.8B, 2B, 8B and 31B. The company claims world-class translation performance and says even the smallest model outperforms Google's TranslateGemma and Tencent's Hy-MT2.
Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune that cuts propaganda rate from 22% to 0.9%
Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to remove Chinese state propaganda from its answers. According to Nous, the resulting Arete-27B model drops the propaganda rate from 22 percent to 0.9 percent while keeping general capabilities intact. Weights, test set and training data are public under Apache 2.0.
PrunaAI releases P-H3-Image: 8 MP image generation and editing in 8 steps, built on MiniMax-H3
PrunaAI has released P-H3-Image, an image model based on MiniMax-H3 that renders 8-megapixel images in 8 steps with no classifier-free guidance. It handles text-to-image and editing with up to 8 input images, and bf16 inference needs about 101 GB of GPU memory.
Comments
Loading...