CPU-only speech models tested for a home server setup
Running at home without a is possible, but the best choice depends on whether sound quality, response speed, or small size matters most. Three models were tested on a 4-core Intel Xeon system with 15.6 gigabytes of memory and no GPU, across 150 runs. For the best sound quality, Kokoro 82M with the ONNX build is the strongest choice.
It produced audio at about 1.8 times speed on this CPU, while the PyTorch build was slower at about 1.3 times speed with nearly the same quality score. For voice or chatbots where quick replies matter more, Supertonic 3 at the 5-step setting is practical. It ran at about 3.2 times speed and scored 4.37 for quality, but its OpenRAIL-M license has commercial limits that need checking before public or business use.
For the smallest setup, Inflect-Nano-v1 is very fast and light, with 4.6 million and about 7.3 times speed, but the voice can sound buzzy and robotic.
Key points
- The test used a 4-core CPU, 15.6 gigabytes of memory, and no GPU.
- Kokoro 82M ONNX is the best fit when voice quality matters most.
- The PyTorch version gives nearly the same quality but runs slower.
- Supertonic 3 at 5 steps is better when fast replies are the priority.
- Inflect-Nano-v1 is tiny and fast, but the voice may sound robotic.