Open TTS Leaderboard measures different parts of speech quality
Intelligibility, speaker similarity and playback delay offer distinct evidence; none replaces listening for naturalness and expression.
Accurate words and natural delivery are different properties. The Open TTS Leaderboard makes that distinction explicit by evaluating intelligibility, reference-speaker similarity and generation performance separately.
Its intelligibility measure compares the requested text with an automatic transcription of the generated speech. Speaker similarity compares representations of the generated voice and a reference recording. The authors caution that neither directly measures expressiveness, naturalness or listener preference.
The streaming view addresses another question: how long until playback can begin? For a streaming model, that means waiting for an initial audio chunk. For a model without streaming, the complete utterance must be generated first. Batched offline throughput measures a different workload.
I would choose the target language and cloning requirement before comparing models, then listen to the samples in the Listen tab. The default English ranking is not evidence of equivalent performance elsewhere. Hardware also belongs in any interpretation of the latency results.
What remains uncertain is how closely these automatic measurements track the preferences of listeners hearing your application's particular content.
Source: Eric Bezzam and colleagues, Hugging Face, Open TTS Leaderboard introduction, published September 30, 2026.