IndexTTS-2.5 · Saudi emotional LoRA

IndexTTS-2.5 with a LoRA on its GPT, fine-tuned on Saudi / Najdi emotional speech Language defaults to AR. Now serving run 7, for a listening test: trained on 11.5k clips from 558 voices in all 8 emotions. On 12 voices it never heard, our automatic judge rates its emotion right 53% of the time, against 63% for the base model, so it is not the recommended model; v2 is. Please listen and tell us whether the emotion sounds weaker or stronger than that. The strongest control is an emotion reference clip of Saudi speech in the emotion you want. Base model for comparison: IndexTeam/IndexTTS-2.5-Demo.