Hugging Face launches Open TTS Leaderboard to rank 8,000 voice models

Hugging Face released the Open TTS Leaderboard on September 30, 2026. This new tool ranks over 8,000 open-source text-to-speech models. The goal is to fix the slow pace of current evaluation methods.

Most existing leaderboards rely on human voting. Users listen to two voices and pick the better one. This process takes weeks to collect enough data. It also makes it hard to include open-source models that require hosting.

The new leaderboard uses objective metrics instead. It measures intelligibility by checking how many words the AI mispronounces. It uses the Qwen3 ASR model to transcribe the audio and compare it to the original text.

Speed is another key metric. The system measures how fast a model generates audio on an H200 GPU. It also tracks the time to first audio for streaming applications. This helps developers find models that respond quickly.

For voice cloning, the leaderboard checks speaker similarity. It compares the voice of the generated audio to a reference clip. This ensures the AI sounds like the intended person, not just a generic robot.

The platform supports multiple languages. English performance does not always predict success in other tongues. Users can toggle between languages to see how models handle different phonetics. Chinese, Japanese, and Korean use character error rates for accuracy.

A "Listen" tab allows users to hear the actual outputs. This bridges the gap between numbers and real-world experience. Users can vote on their preferences, adding a human element to the objective data.

The team plans to open-source the evaluation scripts. This will let the community suggest new datasets or metrics. The leaderboard aims to be shaped by user feedback rather than just technical benchmarks.

This tool does not measure naturalness or emotional expression. It focuses on clarity, speed, and identity preservation. Human preference remains the ultimate judge for artistic quality.

Will objective metrics change how we choose voice assistants? The leaderboard suggests that speed and accuracy are just as important as style.