File size: 8,713 Bytes
7bd0f2f
 
 
 
 
 
 
 
 
 
 
 
 
 
b2804d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9e8c199
2ddf9be
b2804d8
 
 
9e8c199
b2804d8
 
 
 
 
 
9e8c199
b2804d8
 
 
 
 
9e8c199
b2804d8
 
 
 
 
 
 
 
 
 
0d700fc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a974e2a
 
0d700fc
 
a974e2a
0d700fc
 
b2804d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1d7bf00
b2804d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e56c389
 
 
 
 
 
 
 
 
 
 
 
 
b2804d8
 
 
 
b49df67
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
---
license: other
language:
- en
- es
- fr
- de
- ru
- zh
metrics:
- roc_auc
pipeline_tag: voice-activity-detection
library_name: nemo
---
# 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. <br>
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.) <br>

This model is ready for commercial use. <br>

### License/Terms of Use
GOVERNING TERMS: Your use of this model is governed by the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license).


Deployment Geography: Global <br>

Use Case: Developers, speech processing engineers, and AI researchers will use it as the first step for other speech processing models. <br>


## 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.](https://arxiv.org/abs/2010.13886)  <br>
[2] [NVIDIA NeMo Toolkit](https://github.com/NVIDIA/NeMo)
<br> 

## Model Architecture:

**Architecture Type:**  Convolutional Neural Network (CNN) <br>
**Network Architecture:** MarbleNet <br>

**This model has 91.5K of model parameters** <br>

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

## Output: <br>
**Output Type(s):** Sequence of speech probabilities for each 20 millisecond frame <br>
**Output Format:** Float Array <br>
**Output Parameters:** 1D <br>
**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). <br>

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
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

```python
import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.EncDecFrameClassificationModel.from_pretrained(model_name="frame_vad_multilingual_marblenet_v2.0.nemo")
```

### Perform VAD Inference

```bash
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=<Path to .nemo file from which model should be instantiated> \
  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-2.0.0 <br> 

**Supported Hardware Microarchitecture Compatibility:** <br>
* [NVIDIA Ampere] <br>
* [NVIDIA Blackwell] <br>
* [NVIDIA Jetson]  <br>
* [NVIDIA Hopper] <br>
* [NVIDIA Lovelace] <br>
* [NVIDIA Pascal] <br>
* [NVIDIA Turing] <br>
* [NVIDIA Volta] <br>

## Preferred/Supported Operating System(s):
* [Linux] <br>

## Model Version(s):
Frame-VAD Multilingual MarbleNet v2.0  <br>

## Training, Testing, and Evaluation Datasets:

### Training Dataset:
**Link:**  
1. [ICSI (en)](https://groups.inf.ed.ac.uk/ami/icsi/download/)
2. [AMI (en)](https://groups.inf.ed.ac.uk/ami/corpus/)
3. [MLS (fr, es)](https://www.openslr.org/94/)
4. [MCV7 (de, ru)](https://commonvoice.mozilla.org/en/datasets)
5. [RULS (ru)](https://www.openslr.org/96/)
6. [SOVA (ru)](https://github.com/sovaai/sova-dataset)
7. [Aishell2 (zh)](https://www.aishelltech.com/)
8. [Librispeech (en)](https://www.openslr.org/12)
9. [Fisher (en)](https://www.ldc.upenn.edu/)
10. [MUSAN (noise)](https://www.openslr.org/17/)
11. [Freesound (noise)](https://freesound.org/)
12. [Vocalsound (noise)](https://github.com/YuanGongND/vocalsound)
13. [Ichbi (noise)](https://bhichallenge.med.auth.gr/ICBHI_2017_Challenge)
14. [Coswara (noise)](https://github.com/iiscleap/Coswara-Data) <br>

Data Collection Method by dataset:  <br>
* Hybrid: Human, Annotated, Synthetic <br>


Labeling Method by dataset:  <br>
* Hybrid: Human, Annotated, Synthetic <br>

**Properties:**
2600 hours of real-world data, 1000 hours of synthetic data, and 330 hours of noise data
<br>


### Testing Dataset:

**Link:** 
1. [Freesound (noise)](https://freesound.org/)
2. [MUSAN (noise)](https://www.openslr.org/17/)
3. [Librispeech (en)](https://www.openslr.org/12)
4. [Fisher (en)](https://www.ldc.upenn.edu/)
5. [MLS (fr, es)](https://www.openslr.org/94/)
6. [MCV7 (de, ru)](https://commonvoice.mozilla.org/en/datasets)
7. [AMI (en)](https://groups.inf.ed.ac.uk/ami/corpus/)
8. [Aishell2 (zh)](https://www.aishelltech.com/)
9. [CH109 (en)](https://catalog.ldc.upenn.edu/LDC97S42)  <br>

Data Collection Method by dataset:  <br>
* Hybrid: Human, Annotated <br>

Labeling Method by dataset:  <br>
* Hybrid: Human, Annotated <br>


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


### Evaluation Dataset:

**Link:** 
1. [VoxConverse-test](https://github.com/joonson/voxconverse/tree/master)
2. [VoxConverse-dev](https://github.com/joonson/voxconverse/tree/master)
3. [AMI-test](https://github.com/BUTSpeechFIT/AMI-diarization-setup/tree/main/only_words/rttms)
4. [Earnings21](https://github.com/revdotcom/speech-datasets/tree/main/earnings21)
5. [AISHELL4-test](https://www.openslr.org/111/)
6. [CH109](https://catalog.ldc.upenn.edu/LDC97S42)
7. [AVA-SPEECH](https://github.com/rafaelgreca/ava-speech-downloader) <br>

Data Collection Method by dataset:  <br>
* Hybrid: Human, Annotated <br>

Labeling Method by dataset:  <br>
* Hybrid: Human, Annotated <br>

**Properties:** 
Around 182 hours of multilingual (Chinese, English) audio data <br>


# Inference:
**Engine:** NVIDIA NeMo <br>
**Test Hardware:** <br>  
* RTX 5000 <br>
* A100 <br>
* V100  <br>

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

| Eval Dataset     |  ROC-AUC |
|------------------|----------|
| VoxConverse-test |   96.65  |
| VoxConverse-dev  |   97.59  |
| AMI-test         |   96.25  |
| Earnings21       |   97.11  |
| AISHELL4-test    |   92.27  |
| CH109            |   94.44  |
| AVA-SPEECH       |   95.26  |


## 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](https://huggingface.co/nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0/blob/main/explainability.md), [Bias](https://huggingface.co/nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0/blob/main/bias.md), [Safety & Security](https://huggingface.co/nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0/blob/main/safety.md), and [Privacy](https://huggingface.co/nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0/blob/main/privacy.md) Subcards [here](https://huggingface.co/nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0/tree/main).  Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).