~/TEXT TO SPEE/tiny-voca-a-fully-open-4-5m-parameter-tts-running-70x-real

tiny-voca: A Fully Open 4.5M-Parameter TTS Running 70x Real-Time on CPU

A developer has released tiny-voca, an open-source text-to-speech model with only 4.5 million parameters that can run directly in the browser and achieve speeds 70 times faster than real-time on a CPU. The project provides complete transparency by open-sourcing its code, weights, custom phonemizer, and training dataset. This ultra-lightweight model enables efficient edge AI and client-side audio generation without relying on server-side infrastructure or heavy neural vocoders. By making all components including the dataset fully open, it provides a valuable resource for researchers and developers working on resource-constrained speech synthesis. Weighing only 18 MB, tiny-voca avoids a neural vocoder by utilizing a small source-filter synth to generate waveforms, supporting PyTorch, ONNX, and browser execution. However, it currently supports only English with two voices, lacks voice cloning, and can mispronounce foreign names or odd spellings.

## BACKGROUND

A neural vocoder is a machine learning model that converts intermediate speech representations like mel-spectrograms into audible time-domain waveforms. A phonemizer is a tool that converts written text into its constituent phonemes, which are the distinct units of sound that differentiate one word from another in a language.

## REFERENCES

## KEYWORDS

#Text-to-Speech#Open Source#Machine Learning#Edge AI#Audio Generation

$ subscribe --daily