Please verify that you are using Indic-Mio checkpoint.

yes https://huggingface.co/SPRINGLab/Indic-Mio this one

i re-tried this f5 one it works this time...but i m unable to add any emotion to the text every thing is said in a regular tone

use this code:
import os
import sys
import shutil
from pathlib import Path
import numpy as np
import torch

--- 1. DEFINE YOUR LOCAL PATHS ---

ENGINE_BASE = Path(r"D:\TTS_Engine")
CKPT_BASE = ENGINE_BASE / "ckpts"
F5_DIR = CKPT_BASE / "f5_hindi"
VOCOS_DIR = CKPT_BASE / "vocos-mel-24khz"
FFMPEG_BIN = ENGINE_BASE / "ffmpeg" / "bin"

if FFMPEG_BIN.exists():
os.environ["PATH"] += os.pathsep + str(FFMPEG_BIN)

--- 2. FORCE HINDI VOCABULARY ---

default_vocab_path = ENGINE_BASE / "venv" / "Lib" / "site-packages" / "f5_tts" / "infer" / "examples" / "vocab.txt"
hindi_vocab_path = F5_DIR / "vocab.txt"

if hindi_vocab_path.exists() and default_vocab_path.exists():
shutil.copy(hindi_vocab_path, default_vocab_path)

--- 3. MONKEYPATCH HUGGINGFACE ---

import huggingface_hub
def mocked_hf_download(repo_id, filename, **kwargs):
if "vocos" in repo_id:
return str(VOCOS_DIR / filename)
if "F5-TTS" in repo_id or "f5tts" in repo_id or "SPRINGLab" in repo_id:
if "safetensors" in filename or "model.safetensors" in filename:
return str(F5_DIR / "model_2500000.safetensors")
if "vocab.txt" in filename:
return str(F5_DIR / "vocab.txt")
return str(F5_DIR / filename)
return ""
huggingface_hub.hf_hub_download = mocked_hf_download

--- 4. DISABLE WHISPER ---

import f5_tts.infer.utils_infer as utils_infer
utils_infer.initialize_asr_pipeline = lambda *args, **kwargs: print("ASR Disabled")
utils_infer.transcribe = lambda *args, **kwargs: ""

--- 5. TEXT FILTER BYPASS ---

def bypassed_pinyin(texts, **kwargs):
return texts
utils_infer.convert_char_to_pinyin = bypassed_pinyin

--- 6. FORCE HINDI ARCHITECTURE (Fixes the RoPE Scramble) ---

original_load_model = utils_infer.load_model
def mocked_load_model(model_cls, model_cfg, ckpt_path, **kwargs):
# Added pe_attn_head=1 & checkpoint_activations=False
# Without these, the attention mechanism scrambles the audio into static!
model_cfg = dict(
dim=768,
depth=18,
heads=12,
ff_mult=2,
text_dim=512,
text_mask_padding=False,
conv_layers=4,
pe_attn_head=1,
checkpoint_activations=False
)
ckpt_path = str(F5_DIR / "model_2500000.safetensors")
kwargs["vocab_file"] = str(F5_DIR / "vocab.txt")
return original_load_model(model_cls, model_cfg, ckpt_path, **kwargs)
utils_infer.load_model = mocked_load_model

--- 7. SAFE INFER (Fixes the CFG Burn) ---

original_infer_process = utils_infer.infer_process
def safe_infer_process(*args, **kwargs):
# FORCE CFG to 1.0 to prevent the ODE solver from amplifying noise
kwargs['cfg_strength'] = 1.0

wave, sr, spec = original_infer_process(*args, **kwargs)

if isinstance(wave, torch.Tensor):
    wave = wave.squeeze().cpu().numpy()
    
if np.isnan(wave).any():
    wave = np.nan_to_num(wave)
    
max_amp = np.max(np.abs(wave))
if max_amp > 1.0 or max_amp == 0.0:
    if max_amp > 0:
        wave = (wave / max_amp) * 0.95 
        
# Force float32 to prevent Gradio wrapping
wave = wave.astype(np.float32)
    
return wave, sr, spec

utils_infer.infer_process = safe_infer_process

--- 8. LAUNCH ---

os.environ["CUDA_VISIBLE_DEVICES"] = "0"
os.environ["HF_HUB_OFFLINE"] = "1"
os.environ["PYTHONHTTPSVERIFY"] = "0"

if name == "main":
from f5_tts.infer.infer_gradio import main
print("\n" + "="*50)
print("RTX 6000 Ada: F5-TTS FINAL NOISE CURE")
print("Fix 1: pe_attn_head=1 applied (Time Alignment Restored)")
print("Fix 2: CFG forced to 1.0 (ODE Explosion Prevented)")
print("="*50 + "\n")
main()