Public TTS benchmark by Podonos
Monthly view of synthetic voice quality from open-source and commercial TTS outputs, scored by the Podonos API. New snapshot each month.
Every model is tested on 50 held-out sentences spanning varied phonemic content and prosodic patterns.
| Model | Overall MOS | Naturalness | Prosody | Pronunciation | Samples |
|---|---|---|---|---|---|
| StyleTTS2-Base |
4.47
|
4.52 | 4.38 | 4.51 | 50 |
| MATCHA-TTS |
4.19
|
4.21 | 4.15 | 4.22 | 50 |
| VITS-LJSpeech |
3.97
|
4.02 | 3.88 | 4.01 | 50 |
| FastSpeech2-EN |
3.61
|
3.58 | 3.55 | 3.70 | 50 |
| SoundStorm-ref |
3.44
|
3.51 | 3.39 | 3.42 | 50 |
| YourTTS-Multi |
3.28
|
3.31 | 3.22 | 3.31 | 50 |
Benchmark method
We test each model on 50 held-out test sentences. Reported scores are mean MOS estimates across 50 samples, with a 95% confidence interval. Sources are publicly available checkpoints or API outputs as of the listed evaluation date. Podonos runs this benchmark and receives no compensation from any TTS vendor for inclusion.
Podonos API scores publicly available model checkpoints. Sentences come from the Harvard Sentence List (public domain). Last updated: June 2026.
Score your model
Missing your architecture? Send outputs through the API to compare them with this benchmark.