One model. Every supported method.
Author libraries, official defaults, and separate results for each generation method. Full runs use all 1,088 English Seed-TTS-Eval texts. Voice cloning uses each target’s published reference audio and transcript.
Coverage & results
Pending and failed installations are visible. Only completed, verified and published evaluations enter the chart. A method candidate remains under review until its checkpoint and native API are verified.
| Model / checkpoint | Method | Status | Audio | WER ↓ | CER ↓ | DNSMOS ↑ | UTMOS22 ↑ | WavLM SIM ↑ | RTF ↓ | Outputs |
|---|
Samples
What the scores mean
WER / CER: Whisper-large-v3 (faster-whisper, CUDA FP16, English, beam 5, temperature 0, no VAD or previous-text conditioning). Both reference and transcript pass through Whisper’s English text normalizer before corpus edit counts are calculated. Empty generations retain a 100% deletion penalty.
DNSMOS / UTMOS22: predicted quality, not human listening-test MOS. DNSMOS uses Microsoft’s SIG/BAK/OVRL and P.808 models; UTMOS22 uses the author’s strong learner checkpoint, not the full ensemble.
SIM: cosine similarity from WavLM-large with a fine-tuned speaker-verification head. It compares generated speech with the actual conditioning reference. Methods without reference speaker audio show N/A. Reference speech is excluded from the generated audio being scored.
Defaults & timing: one predeclared seed per target; no selection of the best seed or sampling result. Distinct native generation methods get separate rows. RTF includes per-request reference preparation; model loading and a separate warmup are excluded. TTFA is measured only when the native API yields audio chunks. Training, input/output file formats and duplicate batching wrappers are not separate generation methods.