NeuTTS for Unity (Open-Phonemizer Edition)

High-performance, on-device Text-to-Speech implementation for Unity. This version replaces espeak with an ONNX-based phonemizer pipeline and uses NeuCodec for high-fidelity audio reconstruction.

🚀 Key Features

• Engine: English-only synthesis.
• Inference: Fully local using GGUF for the backbone and ONNX for the codec and phonemizer.

---

📦 1. Required Files

Store these files in a permanent directory on your local machine (e.g., C:/Models/NeuTTS/ or /Users/Shared/Models/).

A. Synthesis & Codec (Neuphonic)

• Backbone: A Neuphonic GGUF model (e.g., neutts-air-q4.gguf).
• Codec Decoder: decoder_model.onnx (NeuCodec ONNX model).

B. Phonemizer (lookbe/open-phonemizer-onnx)

• model.onnx
• tokenizer.json
• phoneme_dict.json

C. Voice Reference

• sample.json (Reference voice codes).
• sample_transcript.txt (Matching text for the reference voice).

---

🛠 2. Preparation: Reference Conversion

Unity cannot read PyTorch .pt files directly. Use this Python script to convert your reference voice codes to plain JSON before use:

import torch
import json

def convert_pt_to_json(input_file, output_file):
    # Load the torch tensor
    data = torch.load(input_file, map_location='cpu')
    
    # Extract codes and convert to list for JSON compatibility
    if isinstance(data, torch.Tensor):
        data = data.detach().numpy().tolist()
    
    with open(output_file, 'w') as f:
        json.dump({"codes": data}, f)
    print(f"Success: {output_file} created.")