📐 The 🗣️ Open TTS Leaderboard benchmarks open-source text-to-speech models on intelligibility (WER), speed (RTFx, TTFA), naturalness (UTMOS for English, KOMOS for Korean), and speaker similarity (SIM).
Models are scored across four multilingual eval sets (Seed-TTS eval, CV3-Eval, MiniMax's TTS-Multilingual-Test-Set and Neyshekar) for the language it supports, both without voice cloning (the model's own default voice) and — where the model supports it — with voice cloning from a reference clip.
⚠️ TO REQUEST A MODEL/DATASET/METRIC: open a pull request or issue on the GitHub repo! See here for adding a new model.
🏅 Leaderboard
Averages across the selected languages and datasets. A model is ranked, with a headline average, only if it has results for every selected language on every selected dataset, so all ranked models are compared on the same set.
If batch inference is supported, the model is evaluated with batch size 32; otherwise, it is evaluated at batch size 1. See the 🗣️ Streaming tab for single-request, streaming-mode latency instead.
Evals are run with HF jobs, flavor h200 (unless noted otherwise): 1x H200 (141 GB), 23 vCPU, 256 GB RAM, 3000 GB storage
See the 🤗 About tab for metric definitions (WER, RTFx, SIM, TTFA, UTMOS, KOMOS).
🏆 Top models
📊 Pareto frontiers
Last updated on 13 November 2026
Pick a past snapshot to see the leaderboard as it stood then.
Versioning follows semver: MINOR bumps when datasets, normalization, or a UI/feature changes (what's measured or how); PATCH bumps for model-only additions/removals or UI bugfixes/features. An entry that adds models alongside a dataset/feature change still counts as MINOR.