gradium/phonon-leaderboard-clone-voices

Phonon reference voices for the Open TTS Leaderboard

Phonon's voice for every reference clip in the Open TTS Leaderboard's voice-cloning evaluation, so the leaderboard can evaluate Phonon's cloning without a cloning API. The released Phonon checkpoints do not include the speaker encoder that turns audio into a voice, so the voices are computed here once. phonon/run_eval.py in huggingface/open-tts-leaderboard reads them with --voice_clone.

Files

• voices/<sha256>.safetensors: one voice per distinct reference clip, a speaker_wavs tensor of shape [1, 512, frames] (speaker-encoder latents at 12.5 frames per second). Each released Phonon checkpoint applies its own speaker projection to it, as it does for its built-in voices.
• index.json: {"<dataset>/<config>/<split>": {"<row id>": "<sha256>"}} for Seed-TTS eval tts/en, CV3-Eval zero_shot/en, fr, de and es, and the MiniMax eval tts/en, fr, de, es and pt.

A voice is named by the SHA-256 of its reference clip's file bytes, as stored in the prompt_audio column of the datasets the leaderboard's prepare_seed_tts_eval.py, prepare_cv3_eval.py and prepare_minimax_eval.py build. The runner checks every row's clip against this hash before using its voice.

How the voices were made

Each clip was decoded, loudness-normalized and resampled to 24 kHz. Clips shorter than 8 seconds were repeated until they reached 8 seconds, and every clip was cut at 10 seconds, then encoded with Phonon's speaker encoder.

Sources: Seed-TTS eval (BytedanceSpeech/seed-tts-eval), CV3-Eval (QwenAudio/CV3-Eval at 68b7682a410d048308d1a12f2fbc266770cd23d5), and the MiniMax TTS Multilingual Test Set (MiniMaxAI/TTS-Multilingual-Test-Set, CC BY-SA 4.0, prompt clips from Mozilla Common Voice) through its leaderboard copy hf-audio/minimax_eval at ec62680f994782fb314e781372c64d77aa592aa0.

Use

These voices are derived from the reference audio of those test sets and are shared only to evaluate Phonon on the Open TTS Leaderboard. Access is granted on request.