Voice Safety Classifier v2
Browse files- README.md +56 -1
- config.json +141 -0
- inference.py +111 -0
- model.safetensors +3 -0
- requirements.txt +4 -0
README.md
CHANGED
@@ -10,4 +10,59 @@ language:
|
|
10 |
- ko
|
11 |
- ja
|
12 |
library_name: transformers
|
13 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
- ko
|
11 |
- ja
|
12 |
library_name: transformers
|
13 |
+
---
|
14 |
+
## Model Description
|
15 |
+
We present a voice-safety classification model that can be used for voice-toxicity detection and classification.
|
16 |
+
The model has been distilled into the [WavLM](https://arxiv.org/abs/2110.13900) architecture from a larger teacher model.
|
17 |
+
All the model training has been conducted with Roblox internal voice chat datasets,
|
18 |
+
using both machine and human-labeled data, with over 100k hours of training data in total.
|
19 |
+
We have also published a blog post about this work.
|
20 |
+
|
21 |
+
The model supports eight languages: English, Spanish, German, French, Portuguese, Italian, Korean, and Japanese.
|
22 |
+
It classifies the input audio into six toxicity classes in a multilabel fashion. The class labels are as follows:
|
23 |
+
`Discrimination`, `Harassment`, `Sexual`, `IllegalAndRegulated`, `DatingAndRomantic`, and `Profanity`.
|
24 |
+
Please refer to [Roblox Community Standards](https://en.help.roblox.com/hc/en-us/articles/203313410-Roblox-Community-Standards)
|
25 |
+
for a detailed explanation on the policy, which has been used for labeling the datasets.
|
26 |
+
The model outputs have been calibrated for the Roblox voice chat environment,
|
27 |
+
so that the class scores after a sigmoid can be interpreted as probabilities.
|
28 |
+
|
29 |
+
The classifier expects 16kHz audio segments as input. Ideal segment length is 15 seconds,
|
30 |
+
but the classifier can operate on shorter segments as well. The prediction accuracy may degrade
|
31 |
+
for longer segments.
|
32 |
+
|
33 |
+
The table below displays evaluation precision and recall for each of the supported languages,
|
34 |
+
as calculated over internal language-specific held-out datasets, which resemble the Roblox voice chat traffic.
|
35 |
+
The operating thresholds for each of the categories were kept equal per language, and optimized
|
36 |
+
to achieve a false positive rate of 1%. The classifier was then evaluated as a binary classifier,
|
37 |
+
tagging the audio as positive if any of the heads exceeded the threshold.
|
38 |
+
|
39 |
+
|Language|Precision|Recall|
|
40 |
+
|---|---|---|
|
41 |
+
|English |63.9%|58.2%|
|
42 |
+
|Spanish |76.1%|63.2%|
|
43 |
+
|German |69.9%|74.1%|
|
44 |
+
|French |70.3%|69.8%|
|
45 |
+
|Portuguese|85.4%|58.0%|
|
46 |
+
|Italian |86.6%|52.4%|
|
47 |
+
|Korean |78.0%|64.6%|
|
48 |
+
|Japanese |56.7%|57.7%|
|
49 |
+
|
50 |
+
Compared to the v1 voice safety classifier, the v2 model
|
51 |
+
expands the support from English to 7 additional languages,
|
52 |
+
as well as significantly improving the classification accuracy.
|
53 |
+
With the 1% false positive rate as above, the binary recall for English is improved by 92%.
|
54 |
+
|
55 |
+
|
56 |
+
## Usage
|
57 |
+
The dependencies for the inference file can be installed as follows:
|
58 |
+
```
|
59 |
+
pip install -r requirements.txt
|
60 |
+
```
|
61 |
+
The provided Python file demonstrates how to use the classifier with arbitrary 16kHz audio input.
|
62 |
+
To run the inference, please run the following command:
|
63 |
+
```
|
64 |
+
python inference.py --audio_file <your audio file path> --model_path <path to Huggingface model>
|
65 |
+
```
|
66 |
+
You can download the model weights from the model releases page [here](https://github.com/Roblox/voice-safety-classifier/releases/tag/vs-classifier-v2),
|
67 |
+
or from HuggingFace under [`roblox/voice-safety-classifier-v2`](https://huggingface.co/Roblox/voice-safety-classifier-v2).
|
68 |
+
If `model_path` isn’t specified, the model will be loaded directly from HuggingFace.
|
config.json
ADDED
@@ -0,0 +1,141 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "patrickvonplaten/wavlm-libri-clean-100h-base-plus",
|
3 |
+
"activation_dropout": 0.0,
|
4 |
+
"adapter_kernel_size": 3,
|
5 |
+
"adapter_stride": 2,
|
6 |
+
"add_adapter": false,
|
7 |
+
"apply_spec_augment": true,
|
8 |
+
"architectures": [
|
9 |
+
"WavLMForSequenceClassification"
|
10 |
+
],
|
11 |
+
"attention_dropout": 0.0,
|
12 |
+
"bos_token_id": 1,
|
13 |
+
"classifier_proj_size": 256,
|
14 |
+
"codevector_dim": 256,
|
15 |
+
"contrastive_logits_temperature": 0.1,
|
16 |
+
"conv_bias": false,
|
17 |
+
"conv_dim": [
|
18 |
+
512,
|
19 |
+
512,
|
20 |
+
512,
|
21 |
+
512,
|
22 |
+
512,
|
23 |
+
512,
|
24 |
+
512,
|
25 |
+
512
|
26 |
+
],
|
27 |
+
"conv_kernel": [
|
28 |
+
10,
|
29 |
+
3,
|
30 |
+
3,
|
31 |
+
3,
|
32 |
+
3,
|
33 |
+
2,
|
34 |
+
2,
|
35 |
+
7
|
36 |
+
],
|
37 |
+
"conv_stride": [
|
38 |
+
5,
|
39 |
+
2,
|
40 |
+
2,
|
41 |
+
2,
|
42 |
+
2,
|
43 |
+
2,
|
44 |
+
2,
|
45 |
+
2
|
46 |
+
],
|
47 |
+
"ctc_loss_reduction": "mean",
|
48 |
+
"ctc_zero_infinity": false,
|
49 |
+
"diversity_loss_weight": 0.1,
|
50 |
+
"do_stable_layer_norm": false,
|
51 |
+
"eos_token_id": 2,
|
52 |
+
"feat_extract_activation": "gelu",
|
53 |
+
"feat_extract_norm": "group",
|
54 |
+
"feat_proj_dropout": 0.0,
|
55 |
+
"feat_quantizer_dropout": 0.0,
|
56 |
+
"final_dropout": 0.0,
|
57 |
+
"freeze_feat_extract_train": true,
|
58 |
+
"hidden_act": "gelu",
|
59 |
+
"hidden_dropout": 0.0,
|
60 |
+
"hidden_size": 1024,
|
61 |
+
"id2label": {
|
62 |
+
"0": "LABEL_0",
|
63 |
+
"1": "LABEL_1",
|
64 |
+
"2": "LABEL_2",
|
65 |
+
"3": "LABEL_3",
|
66 |
+
"4": "LABEL_4",
|
67 |
+
"5": "LABEL_5"
|
68 |
+
},
|
69 |
+
"initializer_range": 0.02,
|
70 |
+
"intermediate_size": 3072,
|
71 |
+
"label2id": {
|
72 |
+
"LABEL_0": 0,
|
73 |
+
"LABEL_1": 1,
|
74 |
+
"LABEL_2": 2,
|
75 |
+
"LABEL_3": 3,
|
76 |
+
"LABEL_4": 4,
|
77 |
+
"LABEL_5": 5
|
78 |
+
},
|
79 |
+
"layer_norm_eps": 1e-05,
|
80 |
+
"layerdrop": 0.0,
|
81 |
+
"mask_channel_length": 10,
|
82 |
+
"mask_channel_min_space": 1,
|
83 |
+
"mask_channel_other": 0.0,
|
84 |
+
"mask_channel_prob": 0.0,
|
85 |
+
"mask_channel_selection": "static",
|
86 |
+
"mask_feature_length": 10,
|
87 |
+
"mask_feature_min_masks": 0,
|
88 |
+
"mask_feature_prob": 0.0,
|
89 |
+
"mask_time_length": 10,
|
90 |
+
"mask_time_min_masks": 2,
|
91 |
+
"mask_time_min_space": 1,
|
92 |
+
"mask_time_other": 0.0,
|
93 |
+
"mask_time_prob": 0.05,
|
94 |
+
"mask_time_selection": "static",
|
95 |
+
"max_bucket_distance": 800,
|
96 |
+
"model_type": "wavlm",
|
97 |
+
"no_mask_channel_overlap": false,
|
98 |
+
"no_mask_time_overlap": false,
|
99 |
+
"num_adapter_layers": 3,
|
100 |
+
"num_attention_heads": 16,
|
101 |
+
"num_buckets": 320,
|
102 |
+
"num_codevector_groups": 2,
|
103 |
+
"num_codevectors_per_group": 320,
|
104 |
+
"num_conv_pos_embedding_groups": 16,
|
105 |
+
"num_conv_pos_embeddings": 128,
|
106 |
+
"num_ctc_classes": 80,
|
107 |
+
"num_feat_extract_layers": 7,
|
108 |
+
"num_hidden_layers": 10,
|
109 |
+
"num_negatives": 100,
|
110 |
+
"output_hidden_size": 768,
|
111 |
+
"pad_token_id": 28,
|
112 |
+
"proj_codevector_dim": 256,
|
113 |
+
"replace_prob": 0.5,
|
114 |
+
"tdnn_dilation": [
|
115 |
+
1,
|
116 |
+
2,
|
117 |
+
3,
|
118 |
+
1,
|
119 |
+
1
|
120 |
+
],
|
121 |
+
"tdnn_dim": [
|
122 |
+
512,
|
123 |
+
512,
|
124 |
+
512,
|
125 |
+
512,
|
126 |
+
1500
|
127 |
+
],
|
128 |
+
"tdnn_kernel": [
|
129 |
+
5,
|
130 |
+
3,
|
131 |
+
3,
|
132 |
+
1,
|
133 |
+
1
|
134 |
+
],
|
135 |
+
"tokenizer_class": "Wav2Vec2CTCTokenizer",
|
136 |
+
"torch_dtype": "float32",
|
137 |
+
"transformers_version": "4.38.2",
|
138 |
+
"use_weighted_layer_sum": false,
|
139 |
+
"vocab_size": 31,
|
140 |
+
"xvector_output_dim": 512
|
141 |
+
}
|
inference.py
ADDED
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright © 2024 Roblox Corporation
|
2 |
+
|
3 |
+
"""
|
4 |
+
This file gives a sample demonstration of how to use the given functions in Python, for the Voice Safety Classifier model.
|
5 |
+
"""
|
6 |
+
|
7 |
+
import torch
|
8 |
+
import librosa
|
9 |
+
import numpy as np
|
10 |
+
import argparse
|
11 |
+
from transformers import WavLMForSequenceClassification
|
12 |
+
|
13 |
+
|
14 |
+
def feature_extract_simple(
|
15 |
+
wav,
|
16 |
+
sr=16_000,
|
17 |
+
win_len=15.0,
|
18 |
+
win_stride=15.0,
|
19 |
+
do_normalize=False,
|
20 |
+
):
|
21 |
+
"""simple feature extraction for wavLM
|
22 |
+
Parameters
|
23 |
+
----------
|
24 |
+
wav : str or array-like
|
25 |
+
path to the wav file, or array-like
|
26 |
+
sr : int, optional
|
27 |
+
sample rate, by default 16_000
|
28 |
+
win_len : float, optional
|
29 |
+
window length, by default 15.0
|
30 |
+
win_stride : float, optional
|
31 |
+
window stride, by default 15.0
|
32 |
+
do_normalize: bool, optional
|
33 |
+
whether to normalize the input, by default False.
|
34 |
+
Returns
|
35 |
+
-------
|
36 |
+
np.ndarray
|
37 |
+
batched input to wavLM
|
38 |
+
"""
|
39 |
+
if type(wav) == str:
|
40 |
+
signal, _ = librosa.core.load(wav, sr=sr)
|
41 |
+
else:
|
42 |
+
try:
|
43 |
+
signal = np.array(wav).squeeze()
|
44 |
+
except Exception as e:
|
45 |
+
print(e)
|
46 |
+
raise RuntimeError
|
47 |
+
batched_input = []
|
48 |
+
stride = int(win_stride * sr)
|
49 |
+
l = int(win_len * sr)
|
50 |
+
if len(signal) / sr > win_len:
|
51 |
+
for i in range(0, len(signal), stride):
|
52 |
+
if i + int(win_len * sr) > len(signal):
|
53 |
+
# padding the last chunk to make it the same length as others
|
54 |
+
chunked = np.pad(signal[i:], (0, l - len(signal[i:])))
|
55 |
+
else:
|
56 |
+
chunked = signal[i : i + l]
|
57 |
+
if do_normalize:
|
58 |
+
chunked = (chunked - np.mean(chunked)) / (np.std(chunked) + 1e-7)
|
59 |
+
batched_input.append(chunked)
|
60 |
+
if i + int(win_len * sr) > len(signal):
|
61 |
+
break
|
62 |
+
else:
|
63 |
+
if do_normalize:
|
64 |
+
signal = (signal - np.mean(signal)) / (np.std(signal) + 1e-7)
|
65 |
+
batched_input.append(signal)
|
66 |
+
return np.stack(batched_input) # [N, T]
|
67 |
+
|
68 |
+
|
69 |
+
def infer(model, inputs):
|
70 |
+
output = model(inputs)
|
71 |
+
probs = torch.sigmoid(torch.Tensor(output.logits))
|
72 |
+
return probs
|
73 |
+
|
74 |
+
|
75 |
+
if __name__ == "__main__":
|
76 |
+
parser = argparse.ArgumentParser()
|
77 |
+
parser.add_argument(
|
78 |
+
"--audio_file",
|
79 |
+
type=str,
|
80 |
+
help="File to run inference",
|
81 |
+
)
|
82 |
+
parser.add_argument(
|
83 |
+
"--model_path",
|
84 |
+
type=str,
|
85 |
+
default="roblox/voice-safety-classifier",
|
86 |
+
help="checkpoint file of model",
|
87 |
+
)
|
88 |
+
args = parser.parse_args()
|
89 |
+
labels_name_list = [
|
90 |
+
"Discrimination",
|
91 |
+
"Harassment",
|
92 |
+
"Sexual",
|
93 |
+
"IllegalAndRegulated",
|
94 |
+
"DatingAndRomantic",
|
95 |
+
"Profanity",
|
96 |
+
]
|
97 |
+
|
98 |
+
# Model is trained on only 16kHz audio
|
99 |
+
audio, _ = librosa.core.load(args.audio_file, sr=16000)
|
100 |
+
input_np = feature_extract_simple(audio, sr=16000)
|
101 |
+
input_pt = torch.Tensor(input_np)
|
102 |
+
model = WavLMForSequenceClassification.from_pretrained(
|
103 |
+
args.model_path, num_labels=len(labels_name_list)
|
104 |
+
)
|
105 |
+
probs = infer(model, input_pt)
|
106 |
+
probs = probs.reshape(-1, 6).detach().tolist()
|
107 |
+
print(f"Probabilities for {args.audio_file}:")
|
108 |
+
for chunk_idx in range(len(probs)):
|
109 |
+
print(f"\nSegment {chunk_idx}:")
|
110 |
+
for label_idx, label in enumerate(labels_name_list):
|
111 |
+
print(f"{label} : {probs[chunk_idx][label_idx]}")
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8311e17845ff80ca00973d0aca5e898c20a3e390f1fd2783a4c227d5b9aec559
|
3 |
+
size 480863848
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch
|
2 |
+
transformers
|
3 |
+
librosa
|
4 |
+
numpy
|