Downloads · 30 days
61
17% of all-time downloads
marblenet-vad-GGUF
cstr/marblenet-vad-GGUF
marblenet-vad-GGUF is a voice activity detection model from cstr. Use it for the voice activity detection task on the model card, and read the license before you ship it in a product. It is set up for ggml. The card lists the license as other.
GGUF conversion of nvidia/FrameVADMultilingualMarbleNetv2.0 for use with CrispStrobe/CrispASR.
All-time downloads
350
Public
Repo size
450 KB
Likes
1
Public
Repo makeup
Click a slice to open those files.
.gguf450 KB · 99%
At a glance
- Task
- Voice Activity Detection
- Library
- ggml
- License
- other
- Access
- Public
- Created
- Apr 27, 2026
- Updated
- Apr 27, 2026
- SHA
- ee5531a3
Base models
Model card
From the Hugging Face model README
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).
- Task
- Voice Activity Detection
- Library
- ggml
- License
- other
- Languages
- en, de, fr, es, ru, zh
- Created
- Apr 27, 2026
- Updated
- Apr 27, 2026