Higgs Audio v3 — Telugu single-speaker voice (ISO-romanized fine-tune)

Expressive single-speaker Telugu (code-switch) TTS, fine-tuned from Higgs Audio v3 (4B). Generates this speaker's voice from text with no reference clip.

Frontend matters: Telugu text must be ISO-15919 romanized with the same frontend.py used in training (Telugu tokenizes to byte-fragments otherwise). generate_speech handles the rest.

import torch, torchaudio
from transformers import AutoModelForCausalLM, AutoTokenizer
from frontend import romanize          # shipped in this repo

repo = "BNarayanaReddy/higgs-telugu-e4-lora-final"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True,
                                             dtype=torch.bfloat16).to("cuda").eval()

text = "హలో, ఈ రోజు ఎలా ఉన్నారు?"
wav = model.generate_speech(romanize(text, "iso"), tok, temperature=0.7, top_p=0.95)
torchaudio.save("out.wav", wav.unsqueeze(0), model.config.sample_rate)

Run #1 of an ISO-romanized single-speaker adaptation. Research use.