PiVo-1.4M: 1.4M Parameter Text-to-Speech Model Runs Locally on CPU

·2026.11.03 00:42

Key point

The model uses a 584,965-parameter acoustic model and an 816,597-parameter vocoder to generate 16 kHz audio without PyTorch or network connections.

PiVo-1.4M is a complete text-to-speech stack with 1,401,562 learned parameters, including the vocoder, designed to answer how few parameters a voice needs.
The system runs entirely locally on a CPU using NumPy and ONNX Runtime, eliminating the need for PyTorch, teacher models, speaker encoders, or network connections at inference time.
Architecture consists of a 584,965-parameter acoustic model that predicts vocoder features and an 816,597-parameter FARGAN-derived vocoder that generates 16 kHz mono audio.
The research model produces very intelligible speech for two English voices, Pi and Vo, though some sentences retain a metallic or robotic texture.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.