gafiatulin/vibevoice-asr-coreai

VibeVoice ASR — speech-to-text (Core AI)

On-device transcription, Qwen2-7B. INT8/4 LM.

Source & export pipeline: github.com/gafiatulin/vibevoice-coreai

On-device performance (M4 Max, Core AI): 11.1× RTF.

⚠️ Beta artifacts. These .aimodel bundles are compiled for macOS 27 / Xcode 27 beta (Core AI). They may need re-export on the GA toolchain. The original weights are Microsoft VibeVoice (see upstream for the model license).

Layout

vibevoice-asr-coreai/
  manifest.json        # role → {variant: path} + recommended flags
  embed_tokens.f16     # host-side embed table
  tokenizer/           # tokenizer files
  lm/lm-embeds-int8.aimodel/
  lm/lm-embeds.aimodel/
  codec/acoustic-encoder.aimodel/
  codec/semantic-encoder.aimodel/
  connector/acoustic-connector.aimodel/
  connector/semantic-connector.aimodel/

Roles

Resolve assets by role via manifest.json (default = recommended variant):

{
  "lm": {
    "default": "lm/lm-embeds.aimodel",
    "int8": "lm/lm-embeds-int8.aimodel"
  },
  "acoustic_encoder": {
    "default": "codec/acoustic-encoder.aimodel"
  },
  "semantic_encoder": {
    "default": "codec/semantic-encoder.aimodel"
  },
  "acoustic_connector": {
    "default": "connector/acoustic-connector.aimodel"
  },
  "semantic_connector": {
    "default": "connector/semantic-connector.aimodel"
  }
}

Recommended flags

{
  "lm": "int4",
  "enc_compute": "gpu",
  "window": 30,
  "overlap": 5
}