notmax123/MamreTTS

🎙️ MamreVoice: State-of-the-Art Hebrew TTS

Project Site | GitHub

MamreVoice is a high-performance, end-to-end Text-to-Speech (TTS) and Voice Conversion system specifically optimized for the Hebrew language. Built on the innovative DiffMamba architecture, it delivers expressive, clear, and lifelike synthetic speech.

✨ Features

• Expressive Hebrew Synthesis: Specialized optimization for Hebrew prosody, rhythm, and naturalness.
• Voice Conversion: High-quality voice cloning and identity preservation.
• Phonetics-Aware: Seamless integration with Phonikud for accurate Hebrew diacritization (Niqqud) and vowelization.
• Efficiency: Designed for low computational overhead, enabling real-time performance on modern Nvidia GPU hardware.

---

🛠️ Technical Overview

Architecture

MamreVoice utilizes DiffMamba, which addresses the computational heavy-lifting of traditional diffusion models by using Mamba blocks. This allows for long-context handling and faster inference without sacrificing audio quality.

Components

1. Text Processing: Advanced normalization and diacritization via Phonikud.
2. BackBone Model: DiffMamba-based Backbone
3. Rest Of The Model: is based on Zonos

---