Speech Synthesis Editing on POLYGLOT-NOUNS curated set of 312 proper nouns (test)
3.1PERtargetFlowEdit
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| FlowEditSteps=∼50, Backbone=F5-TTS2026.06 | 3.1 | 3.22 | 78.6 | 15 | |
| Fine-tuningSteps=500, Learning Rate=10−52026.06 | 8.2 | 4.1 | 74.8 | 20 | |
| LoRARank (r)=16, Steps=5002026.06 | 11.8 | 4.65 | 71.4 | 8 | |
| PromptingPrefix tokens=82026.06 | 18.3 | 5.31 | 69.8 | 5 | |
| LexiconEngine=eSpeak-NG lexicon override2026.06 | 18.7 | 5.61 | 68.3 | — | |
| Zero-shotBackbone=F5-TTS, Parameters=335M2026.06 | 42.5 | 6.82 | 72.1 | — |