license: other language:

• en
• es
• fr
• de
• ru
• zh metrics:
• roc_auc pipeline_tag: voice-activity-detection library_name: nemo tags:
• Multilingual
• MarbleNet
• pytorch
• speech
• audio
• VAD
• onnx
• onnxruntime

Frame-VAD Multilingual MarbleNet v2.0

Description:

Frame-VAD Multilingual MarbleNet v2.0 is a convolutional neural network for voice activity detection (VAD) that serves as the first step for Speech Recognition and Speaker Diarization. It is a frame-based model that outputs a speech probability for each 20 millisecond frame of the input audio. The model has 91.5K parameters, making it lightweight and efficient for real-time applications.
To reduce false positive errors — cases where the model incorrectly detects speech when none is present — the model was trained with white noise and real-word noise perturbations. During training, the volume of audios was also varied. Additionally, the training data includes non-speech audio samples to help the model distinguish between speech and non-speech sounds (such as coughing, laughter, and breathing, etc.)

Key Features

• Lightweight model with only 91.5K parameters
• Robust against false positive errors
• Outputs speech probability for each 20 ms audio frame
• Multilingual support: Chinese, English, French, German, Russian, and Spanish

This model is ready for commercial use.

License/Terms of Use:

GOVERNING TERMS: Your use of this model is governed by the NVIDIA Open Model License Agreement.

Deployment Geography:

Global

Use Case:

Developers, speech processing engineers, and AI researchers will use it as the first step for other speech processing models.

References:

[1] Jia, Fei, Somshubra Majumdar, and Boris Ginsburg. "MarbleNet: Deep 1D Time-Channel Separable Convolutional Neural Network for Voice Activity Detection." ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021.
[2] NVIDIA NeMo Toolkit

Model Architecture:

Architecture Type: Convolutional Neural Network (CNN)
Network Architecture: MarbleNet

This model has 91.5K parameters

Input:

Input Type(s): Audio
Input Format: .wav files
Input Parameters: 1D
Other Properties Related to Input: 16000 Hz Mono-channel Audio, Pre-Processing Not Needed

Output:

Output Type(s): Sequence of speech probabilities for each 20 millisecond frame
Output Format: Float Array
Output Parameters: 1D
Other Properties Related to Output: May need post-processing, such as smoothing, which reduces sudden fluctuations in detected speech probability for more natural transitions, and thresholding, which sets a cutoff value to determine whether a frame contains speech based on probability (e.g., classifying frames above 0.5 as speech and others as silence or noise).

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

How to Use the Model:

To train, fine-tune or play with the model you will need to install NVIDIA NeMo.

pip install -U nemo_toolkit['asr']

The model is available for use in the NeMo toolkit [2], and can be used as a pre-trained checkpoint for inference.

Automatically load the model

import torch
import nemo.collections.asr as nemo_asr
vad_model = nemo_asr.models.EncDecFrameClassificationModel.from_pretrained(model_name="nvidia/frame_vad_multilingual_marblenet_v2.0")

# Move the model to GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
vad_model = vad_model.to(device)
vad_model.eval()

Inference with PyTorch

First, let's get a sample

wget https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav

Then run the following:

import librosa

# Load the audio
input_signal = librosa.load("2086-149220-0033.wav", sr=16000, mono=True)[0]
input_signal = torch.tensor(input_signal).unsqueeze(0).float()
input_signal_length = torch.tensor([input_signal.shape[1]]).long()

# Perform inference
with torch.no_grad():
   torch_outputs = vad_model(
        input_signal=input_signal.to(device),
        input_signal_length=input_signal_length.to(device)
    ).cpu()

Export to ONNX

import onnx 
from nemo.core import typecheck
typecheck.set_typecheck_enabled(False)

vad_model = vad_model.cpu()
ONNX_EXPORT_PATH = "frame_vad_multilingual_marblenet_v2.0.onnx"

# Preprocess input signal
processed_signal, processed_signal_length = vad_model.preprocessor(
    input_signal=input_signal,
    length=input_signal_length
)

# Define input example for ONNX export
inputs = {
    "processed_signal": processed_signal,
    "processed_signal_length": processed_signal_length
}

# Export
torch.onnx.export(
    model=vad_model,
    args=inputs,
    f=ONNX_EXPORT_PATH,
    input_names=list(inputs.keys()),
    output_names=["output"],
    dynamic_axes={
        "processed_signal": {0: "batch_size", 2: "sequence_length"},
        "processed_signal_length": {0: "batch_size"},
        "output": {0: "batch_size", 1: "sequence_length"}
    }
)

Inference with ONNX Runtime

import onnxruntime

# Load the ONNX model
session = onnxruntime.InferenceSession(
    ONNX_EXPORT_PATH, 
    providers=["CPUExecutionProvider"]
)

# Prepare input for ONNX Runtime
ort_inputs = {
    input.name: inputs[input.name].numpy()
    for input in session.get_inputs()
}

# Run inference
onnx_outputs = session.run(None, ort_inputs)[0]

RTTM Output from Frame-Level Speech Predictions

To generate RTTM (Rich Transcription Time Marked) files from audio using the pretrained model:

python <NEMO_ROOT>/examples/asr/speech_classification/frame_vad_infer.py \
  --config-path="../conf/vad" \
  --config-name="frame_vad_infer_postprocess.yaml" \
  vad.model_path="nvidia/frame_vad_multilingual_marblenet_v2.0" \
  vad.parameters.shift_length_in_sec=0.02 \
  prepare_manifest.auto_split=True \
  prepare_manifest.split_duration=7200 \
  input_manifest=<Path of manifest file of evaluation data, where audio files should have unique names> \
  out_manifest_filepath=<Path of output manifest file>

Software Integration:

Runtime Engine(s):

• NeMo-2.3.0
  

Supported Hardware Microarchitecture Compatibility:

• [NVIDIA Ampere]
  
• [NVIDIA Blackwell]
  
• [NVIDIA Jetson]
  
• [NVIDIA Hopper]
  
• [NVIDIA Lovelace]
  
• [NVIDIA Pascal]
  
• [NVIDIA Turing]
  
• [NVIDIA Volta]
  

Preferred/Supported Operating System(s):

• [Linux]
  

Model Version(s):

Frame-VAD Multilingual MarbleNet v2.0

Training, Testing, and Evaluation Datasets:

Training Dataset:

Link:

1. ICSI (en)
2. AMI (en)
3. MLS (fr, es)
4. MCV7 (de, ru)
5. RULS (ru)
6. SOVA (ru)
7. Aishell2 (zh)
8. Librispeech (en)
9. Fisher (en)
10. MUSAN (noise)
11. Freesound (noise)
12. Vocalsound (noise)
13. Ichbi (noise)
14. Coswara (noise)
    

Data Collection Method by dataset:

• Hybrid: Human, Annotated, Synthetic
  

Labeling Method by dataset:

• Hybrid: Human, Annotated, Synthetic
  

Properties: 2600 hours of real-world data, 1000 hours of synthetic data, and 330 hours of noise data

Testing Dataset:

Link:

1. Freesound (noise)
2. MUSAN (noise)
3. Librispeech (en)
4. Fisher (en)
5. MLS (fr, es)
6. MCV7 (de, ru)
7. AMI (en)
8. Aishell2 (zh)
9. CH109 (en)
   

Data Collection Method by dataset:

• Hybrid: Human, Annotated
  

Labeling Method by dataset:

• Hybrid: Human, Annotated
  

Properties: Around 100 hours of multilingual (Chinese, English, French, German, Russian, Spanish) audio data

Evaluation Dataset:

Link:

1. VoxConverse-test
2. VoxConverse-dev
3. AMI-test
4. Earnings21
5. AISHELL4-test
6. CH109
7. AVA-SPEECH
   

Data Collection Method by dataset:

• Hybrid: Human, Annotated
  

Labeling Method by dataset:

• Hybrid: Human, Annotated
  

Properties: Around 182 hours of multilingual (Chinese, English) audio data

Inference:

Engine: NVIDIA NeMo
Test Hardware:

• RTX 5000
  
• A100
  
• V100
  

Performance:

The ROC-AUC performance is listed in the following table. A higher ROC-AUC indicates better performance.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards here.

Please report security vulnerabilities or NVIDIA AI Concerns here.

Bias

Explainability

Privacy

Safety