File size: 3,913 Bytes
3f7a513
 
 
 
51af749
3f7a513
6aeb766
 
 
 
 
3f7a513
8e3d0e5
0dad9f3
 
 
 
6eba34a
 
 
 
 
 
 
 
 
 
 
 
 
 
0dad9f3
9eb0395
0dad9f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ea7d89
0dad9f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6a9abe6
0dad9f3
 
 
 
 
 
 
 
 
 
 
 
6a9abe6
0dad9f3
6a9abe6
0dad9f3
6a9abe6
0dad9f3
6a9abe6
 
8e3d0e5
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
---
language: en
datasets:
- msp-podcast
inference: true
tags:
- speech
- audio
- wav2vec2
- audio-classification
- emotion-recognition
license: cc-by-nc-sa-4.0
pipeline_tag: audio-classification
---

# Model for Dimensional Speech Emotion Recognition based on Wav2vec 2.0

Please note that this model is for research purpose only.
A commercial license for a model
that has been trained on much more data
can be acquired with [audEERING](https://www.audeering.com/products/devaice/).
The model expects a raw audio signal as input,
and outputs predictions for arousal, dominance and valence in a range of approximately 0...1.
In addition,
it provides the pooled states of the last transformer layer.
The model was created by fine-tuning
[Wav2Vec2-Large-Robust](https://huggingface.co/facebook/wav2vec2-large-robust)
on [MSP-Podcast](https://ecs.utdallas.edu/research/researchlabs/msp-lab/MSP-Podcast.html) (v1.7).
The model was pruned from 24 to 12 transformer layers before fine-tuning.
An [ONNX](https://onnx.ai/) export of the model is available from [doi:10.5281/zenodo.6221127](https://zenodo.org/record/6221127).
Further details are given in the associated [paper](https://arxiv.org/abs/2203.07378) and [tutorial](https://github.com/audeering/w2v2-how-to).

# Usage

```python
import numpy as np
import torch
import torch.nn as nn
from transformers import Wav2Vec2Processor
from transformers.models.wav2vec2.modeling_wav2vec2 import (
    Wav2Vec2Model,
    Wav2Vec2PreTrainedModel,
)


class RegressionHead(nn.Module):
    r"""Classification head."""

    def __init__(self, config):

        super().__init__()

        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.dropout = nn.Dropout(config.final_dropout)
        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

    def forward(self, features, **kwargs):

        x = features
        x = self.dropout(x)
        x = self.dense(x)
        x = torch.tanh(x)
        x = self.dropout(x)
        x = self.out_proj(x)

        return x


class EmotionModel(Wav2Vec2PreTrainedModel):
    r"""Speech emotion classifier."""

    def __init__(self, config):

        super().__init__(config)

        self.config = config
        self.wav2vec2 = Wav2Vec2Model(config)
        self.classifier = RegressionHead(config)
        self.init_weights()

    def forward(
            self,
            input_values,
    ):

        outputs = self.wav2vec2(input_values)
        hidden_states = outputs[0]
        hidden_states = torch.mean(hidden_states, dim=1)
        logits = self.classifier(hidden_states)

        return hidden_states, logits



# load model from hub
device = 'cpu'
model_name = 'audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim'
processor = Wav2Vec2Processor.from_pretrained(model_name)
model = EmotionModel.from_pretrained(model_name).to(device)

# dummy signal
sampling_rate = 16000
signal = np.zeros((1, sampling_rate), dtype=np.float32)


def process_func(
    x: np.ndarray,
    sampling_rate: int,
    embeddings: bool = False,
) -> np.ndarray:
    r"""Predict emotions or extract embeddings from raw audio signal."""

    # run through processor to normalize signal
    # always returns a batch, so we just get the first entry
    # then we put it on the device
    y = processor(x, sampling_rate=sampling_rate)
    y = y['input_values'][0]
    y = y.reshape(1, -1)
    y = torch.from_numpy(y).to(device)

    # run through model
    with torch.no_grad():
        y = model(y)[0 if embeddings else 1]

    # convert to numpy
    y = y.detach().cpu().numpy()

    return y


print(process_func(signal, sampling_rate))
#  Arousal    dominance valence
# [[0.5460754  0.6062266  0.40431657]]

print(process_func(signal, sampling_rate, embeddings=True))
# Pooled hidden states of last transformer layer
# [[-0.00752167  0.0065819  -0.00746342 ...  0.00663632  0.00848748
#    0.00599211]]
```