cstr/marblenet-vad-GGUF

MarbleNet VAD -- GGUF

GGUF conversion of nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0 for use with CrispStrobe/CrispASR.

Available variants

File	Size	Notes
marblenet-vad.gguf	439 KB	F32, BatchNorm fused into conv weights

No quantization needed — model is already 91.5K params (439 KB).

Model details

• Architecture: MarbleNet — 1D time-channel separable CNN (6 Jasper blocks: depthwise conv + pointwise conv + BN + ReLU)
• Parameters: 91.5K (smallest VAD model in CrispASR)
• Languages: Chinese, English, French, German, Russian, Spanish
• Input: 80-bin mel spectrogram (16kHz, 512 FFT, 25ms window, 10ms stride)
• Output: per-frame speech probability (20ms per frame)
• Training data: 2,600h real + 1,000h synthetic + 330h noise
• License: NVIDIA Open Model License (commercial use OK)

Benchmark (ROC-AUC from NVIDIA)

Dataset	ROC-AUC
VoxConverse-test	96.65
VoxConverse-dev	97.59
AMI-test	96.25
Earnings21	97.11

Usage with CrispASR

# Auto-download (439 KB)
crispasr --backend whisper -m auto --auto-download --vad -vm marblenet -f audio.wav

# Or with explicit path
crispasr --backend parakeet -m auto --auto-download --vad -vm marblenet-vad.gguf -f audio.wav

CrispASR VAD comparison

VAD Model	Size	Latency	Languages
Silero VAD v5	0.9 MB	~60 ms	Multilingual
MarbleNet	0.4 MB	~30 ms	6 languages
FireRedVAD	2.4 MB	~50 ms	100+ (recommended)
Whisper-VAD-ASMR	22 MB	~1000 ms	Experimental

Conversion

python models/convert-marblenet-vad-to-gguf.py \
  --input nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0 \
  --output marblenet-vad.gguf

BatchNorm layers are fused into convolution weights at convert time (36 tensors from 84 original).