Open TTS Leaderboard Launches Objective Metrics for Multilingual Text-to-Speech and Voice Cloning

A new leaderboard for Open TTS has been introduced to assess text-to-speech models through objective criteria, aiming to reduce fragmentation in TTS assessments. It evaluates intelligibility by calculating word and character error rates with Qwen3 ASR, measures speed using inverse real-time factor and time-to-first-audio on H200 GPU and CPU, and assesses speaker similarity through cosine similarity of WavLM embeddings. The time required for evaluations has been cut down from weeks to mere hours. The top models, based on macro-average WER for English splits of Seed TTS Eval and CV3 Eval, include hexgrad/Kokoro-82M, Supertone/supertonic-3, and fishaudio/s2-pro. Multilingual capabilities feature k2-fsa/OmniVoice and FunAudioLLM/Fun-CosyVoice3-0.5B-2512. This leaderboard prioritizes open-source models and intends to release evaluation scripts. As of September 30, 2026, only 16 out of 92 models on Artificial Analysis are open-weight.