aoiandroid/VibeVoice-Realtime-0.5B-CoreML

VibeVoice-Realtime-0.5B CoreML

CoreML conversion of microsoft/VibeVoice-Realtime-0.5B for iOS 18+.

Model Components

Model	Size	Quantized
TTS LLM (Stateful)	829 MB	208 MB
Diffusion Head	80 MB	20 MB
Acoustic Decoder	656 MB	164 MB

Total: 1,565 MB (original) / 392 MB (quantized 4-bit)

Requirements

• iOS 18+ (Stateful Models for TTS LLM)
• CoreML framework

Files

• vibevoice_tts_llm_quantized.mlpackage - TTS LLM (recommended)
• vibevoice_diffusion_head_quantized.mlpackage - Diffusion Head
• vibevoice_acoustic_decoder_quantized.mlpackage - Acoustic Decoder
• conversion_report.md - Full conversion details
• benchmark_results.json - Performance metrics

Inference Pipeline

1. Tokenize text (Qwen2 tokenizer)
2. TTS LLM -> hidden states
3. Diffusion Head (20 steps) -> acoustic latents
4. Acoustic Decoder -> audio waveform (24kHz)

Source

Converted from microsoft/VibeVoice-Realtime-0.5B.