Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning

The Open TTS Leaderboard compares text-to-speech models with automatic measurements instead of listener votes. It checks how closely a transcription of the generated audio matches the requested text, using the Qwen3 ASR speech-recognition model. It measures how closely the voice resembles a reference voice by comparing speaker embeddings, and it records generation speed. The post says these are proxies: they do not directly measure naturalness, expressiveness or listener preference, and the leaderboard does not replace human preference ranking. The methodology explains each score.

For voice applications that must respond quickly, a separate view ranks models by time to first audio: for streaming models, the wait until the first audio chunk arrives, and for non-streaming models, the wait until the whole utterance is generated. Each model runs one request at a time on the same hardware. Languages can be toggled to rank multilingual performance, and the post notes that one of its evaluation sets only has audio for English and Chinese. A Listen tab lets people compare outputs side by side, which is how voice quality can still be judged by ear. Evaluation and listening options.