f5_tts_loader.py

import torch
import torchaudio
from transformers import AutoTokenizer


class F5TTS:
    def __init__(self, model_name="hynt/F5-TTS-Vietnamese-100h"):
        self.device = torch.device("cpu")

        # Load tokenizer
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)

        # Load model weights
        self.model = torch.load(
            self._download_model(model_name),
            map_location=self.device
        )

        self.model.eval()

    def _download_model(self, repo):
        """Download model file from HuggingFace repo."""
        from huggingface_hub import hf_hub_download
        return hf_hub_download(repo_id=repo, filename="model.safetensors")

    def tts(self, text, sample_rate=22050):
        tokens = self.tokenizer(text, return_tensors="pt")["input_ids"].to(self.device)

        with torch.no_grad():
            audio = self.model.generate(tokens)[0].cpu().numpy()

        return audio, sample_rate