cstr / vibevoice-asr-bitnet-GGUF

VibeVoice-ASR-BitNet GGUF

GGUF conversion of microsoft/VibeVoice-ASR-BitNet for use with CrispASR .

Model Details

Quantization

The LM projection weights are BitNet-trained (ternary {-1, 0, 1}) and stored using ggml's native TQ2_0 format (2.06 bpw). VAE encoder weights use Q8_0 (8-bit). Embedding stays F16. No custom kernels needed; standard ggml mul_mat handles all types transparently.

Usage

CLI

# Auto-download and transcribe
crispasr -m auto --backend vibevoice-bitnet -f audio.wav

# Or with local file
crispasr -m vibevoice-asr-bitnet-tq2.gguf --backend vibevoice -f audio.wav

C API

#include "vibevoice.h"

vibevoice_context_params params = vibevoice_context_default_params();
params.n_threads = 4;
vibevoice_context *ctx = vibevoice_init_from_file("vibevoice-asr-bitnet-tq2.gguf", params);
char *text = vibevoice_transcribe(ctx, samples_24khz, n_samples);
printf("%s\n", text);
free(text);
vibevoice_free(ctx);

Python

import crispasr
s = crispasr.Session("vibevoice-asr-bitnet-tq2.gguf", backend="vibevoice")
result = s.transcribe("audio.wav")
print(result["text"])

Conversion

Converted from the original safetensors using:

python models/convert-vibevoice-bitnet-to-gguf.py \
    --input microsoft/VibeVoice-ASR-BitNet \
    --output vibevoice-asr-bitnet-tq2.gguf

The converter applies Microsoft's ternary quantization ( s = 1/mean(|w|) , round to {-1, 0, 1}) to LM projection weights, then packs them into ggml's TQ2_0 format.

Acknowledgements

• microsoft/VibeVoice for the model architecture and training
• microsoft/VibeASR.cpp for the BitNet quantization approach