CPU speech model test shows the speed and quality tradeoff
Several small models were tested on CPU only, without using a . The models were Kokoro 82M, Supertonic 3, Inflect-Nano-v1, and Kyutai's Pocket TTS. The test machine used an Intel Xeon 8272CL with 4 CPU cores and 15.6GB of memory, with CUDA turned off. Six text lengths were used, from 12 to 1,712 characters, and each setup was timed five times after one warmup run, for 180 total runs.
Audio quality was scored with UTMOS. Supertonic 3 with 2 steps was fastest, with an average RTF of 0.121, but its quality score was low at 1.53. Inflect-Nano-v1 had an RTF of 0.145 and a quality score of 3.48. Supertonic 3 with 5 steps had an RTF of 0.240 and a quality score of 4.32.
Kokoro 82M scored highest on quality, about 4.44 to 4.46, with an RTF around 0.64 to 0.67. Pocket TTS had an RTF of 0.714 and a quality score of 4.10, making it usable on CPU but not the fastest option.
Key points
- The test compared small models on a 4-core CPU with no .
- Supertonic 3 with 2 steps was fastest but had weak audio quality.
- Kokoro 82M had the best quality score, but it was slower.
- Pocket TTS used a different design and reached usable quality on CPU.
- The results are useful when choosing a low-cost speech layer for .