VoxCPM2 ANE

This repository contains the Apple Neural Engine (ANE) optimized CoreML export of the openbmb/VoxCPM2 Text-to-Speech (TTS) model.

These models have been converted to Apple's .mlmodelc / .mlpackage format to leverage hardware acceleration provided by the Apple Neural Engine on Apple Silicon devices. This enables highly efficient and fast on-device text-to-speech generation.

Model Components

The overall voxcpm2 TTS pipeline is broken down into several CoreML sub-modules:

• Language Models (LM): base_lm_multifunction.mlmodelc, residual_lm_fused_multifunction.mlmodelc
• Audio VAE: audio_vae_encoder.mlmodelc, audio_vae_decoder_lf4.mlmodelc
• Feature Encoder: feat_encoder.mlmodelc
• Diffusion Transformer (DiT): locdit_p4_c4.mlmodelc
• Quantization & Projections: fsq_s4.mlmodelc, projections.mlmodelc

Architecture Highlights

Based on the config.json, the model utilizes:

• A base Language Model with 28 layers and 2048 hidden size.
• A Residual LM with 8 layers.
• An Audio VAE supporting up to 48kHz output sample rate.
• A DiT component with 12 layers and 1024 hidden dimension for precise audio modeling.