Methodology

AudioLab turns listener preference into a public ranking. Each session starts with one prompt, generates two voice clips from eligible TTS models, and asks the listener to choose the voice they prefer.

The leaderboard is shaped by aggregate voting behavior rather than editorial selection. Every completed vote contributes to the model ratings.

The app randomly selects two active models, sends both the same text, and presents the generated clips as Voice A and Voice B. Model names stay hidden until the vote is complete.

After a vote, AudioLab reveals the model names and providers, stores the result, and updates the leaderboard.

Ratings use Elo-style pairwise updates. Every model starts from the same baseline, and each winner/loser result moves the two ratings according to the configured K-factor.

Models with limited vote volume are marked as preliminary, and confidence intervals communicate uncertainty while the sample size is still developing.

AudioLab does not store full submitted text in the database by default. Battle records use text hashes and text length, while generated audio is cached in private storage and served through authenticated routes.

Prompts must still be sent to moderation and TTS providers to generate the audio. The service is designed to avoid exposing provider keys, raw model IDs, and private storage paths to the browser.