Aivis-Project/AivisSpeech

What it solves

AivisSpeech is a Japanese text-to-speech (TTS) software designed to generate emotionally rich voices. It provides a user-friendly editor UI based on VOICEVOX, allowing users to easily create synthetic speech from text.

How it works

The software integrates the AivisSpeech Engine, which utilizes the AIVMX (Aivis Voice Model for ONNX) open file format. This format bundles trained models, hyperparameters, style vectors, and speaker metadata into a single file. By using ONNX Runtime, the system achieves fast CPU inference and eliminates dependencies on PyTorch, reducing the installation size.

Who it’s for

It is primarily aimed at content creators and developers who need high-quality Japanese synthetic speech. It supports Windows and macOS (specifically Apple Silicon) users.

Highlights

• ONNX-based Inference: Uses AIVMX files for high-speed CPU performance and a smaller footprint.
• VOICEVOX-based UI: Leverages a familiar editor interface for intuitive voice synthesis.
• AIVM Specification: Supports a standardized open file format for voice models, including metadata and speaker info.
• Architecture Support: Compatible with Style-Bert-VITS2 and Style-Bert-VITS2 (JP-Extra) model architectures.