Four open speech models tested for speed and quality on CPUs

A ed Neo wrote the code, designed the test, built the testing setup, and timed 180 speech on a CPU. The four open models were Kokoro 82M, Supertonic 3, Inflect-Nano, and Kyutai's Pocket TTS, with the results organized into charts and table files.

Pocket TTS copied a voice from only five seconds of audio and worked on a CPU without a GPU or . Kokoro 82M sounded the most natural and generated audio at about 1.5 times real-time speed.

Supertonic 3 reached about four times real-time speed in its quality mode. UTMOS gave Inflect-Nano a reasonable score, but human listening found it buzzy and robotic, showing that the automatic score did not match the perceived quality.

Key points

  • Four open models were compared across 180 timed CPU .
  • Pocket TTS copied a voice from five seconds of audio without a GPU or .
  • Kokoro 82M sounded the most natural and ran at about 1.5 times real-time speed.
  • Supertonic 3 ran at about four times real-time speed in quality mode.
  • Inflect-Nano's UTMOS score was better than its buzzy, robotic sound suggested.
Read original