EvoAwaken-Workshop/VibeVoice-ASR-Streaming-7B-GGUF

VibeVoice-ASR-Streaming-7B GGUF

Unofficial GGUF conversion of `microsoft/VibeVoice-ASR-Streaming-7B` for the native Rust implementation in `Liyulingyue/rust-model-inference`.

The model performs streaming, speaker-attributed automatic speech recognition for Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. It can also use context terms supplied through the Rust CLI's --prompt option.

Files

Both files are required for ASR inference.

The LLM has 28 layers, a hidden size of 3584, a context length of 131072, and a vocabulary size of 152064. The mmproj expects 24 kHz audio and uses a custom vibevoice_asr audio projector.

Compatibility

The validated runtime is rust-model-inference after merged PR #47.

The LLM uses llama.cpp-style Qwen2 tensor names, but the end-to-end ASR model is not compatible with stock llama.cpp. The audio encoder/projector is a VibeVoice-specific mmproj that requires the Rust implementation above.

Rust Usage

git clone https://github.com/Liyulingyue/rust-model-inference.git
cd rust-model-inference

cargo run --release --bin rust-model-inference -- \
  --model /path/to/VibeVoice-ASR-Streaming-7B-Q8_0.gguf \
  --mmproj /path/to/mmproj-VibeVoice-ASR-Streaming-7B-BF16.gguf \
  --audio /path/to/input.wav \
  --temp 0 \
  --max-tokens 128 \
  --threads 8 \
  --out transcription.txt

The current CLI accepts PCM16 WAV input. It downmixes multi-channel input and resamples it to the model's 24 kHz sample rate. --max-tokens is the maximum number of generated tokens per streaming chunk.

Use --prompt to provide names or domain terms as recognition context:

  --prompt "Microsoft,VibeVoice"

For reproducible encoder latents, set:

export VIBEVOICE_ASR_DETERMINISTIC=1

Conversion

The converter reads the original sharded BF16 safetensors through mmap and does not require PyTorch:

python3 tools/vibevoice/convert_vibevoice_asr.py \
  /path/to/VibeVoice-ASR-Streaming-7B \
  --out-dir /path/to/output

The language decoder is quantized to GGML Q8_0. The acoustic and semantic encoders and connectors remain BF16. The unused acoustic tokenizer decoder is not included because ASR does not synthesize audio.

Alignment

The Rust implementation was checked against the original BF16 safetensors on a fixed prompt and audio fixture:

The top-8 token ordering matched exactly:

[58, 151665, 715, 42474, 39379, 32622, 43504, 37073]

The BF16 audio encoder Oracle covered 12 checkpoints with a maximum absolute error of 0.000070.

These are numerical comparisons, not bitwise equality. Q8_0 weight and activation quantization necessarily differs from the original BF16 model. The decoder comparison covers the final sequence row at every layer for one fixed fixture; it does not claim exhaustive equality for every input or internal Q/K/V tensor.

Checksums

255ca05bb3b4f34ab02922da5bc3f4ac6de489552175cc013c1a1db7e69d9514  VibeVoice-ASR-Streaming-7B-Q8_0.gguf
020e256a28f9ccb1e0ad658467545a6a81bd9d3ca2f52cb4ba146712afc1699b  mmproj-VibeVoice-ASR-Streaming-7B-BF16.gguf

Limitations

• —Only the CPU inference path has been validated for this conversion.
• —Quantization can change recognition output relative to the BF16 checkpoint.
• —The Rust CLI currently requires PCM16 WAV files; microphone/WebSocket serving is not included in this model repository.
• —Review the original model card for intended use, evaluation, and responsible use guidance.

License and Attribution

The source model is provided by Microsoft under the MIT License. This repository redistributes converted model weights under the same license. See `LICENSE` and the original model card.