README.md
| 1 | --- |
| 2 | language: en |
| 3 | datasets: |
| 4 | - msp-podcast |
| 5 | inference: true |
| 6 | tags: |
| 7 | - speech |
| 8 | - audio |
| 9 | - wav2vec2 |
| 10 | - audio-classification |
| 11 | - emotion-recognition |
| 12 | license: cc-by-nc-sa-4.0 |
| 13 | pipeline_tag: audio-classification |
| 14 | --- |
| 15 | |
| 16 | # Model for Dimensional Speech Emotion Recognition based on Wav2vec 2.0 |
| 17 | |
| 18 | Please note that this model is for research purpose only. |
| 19 | A commercial license for a model |
| 20 | that has been trained on much more data |
| 21 | can be acquired with [audEERING](https://www.audeering.com/products/devaice/). |
| 22 | The model expects a raw audio signal as input, |
| 23 | and outputs predictions for arousal, dominance and valence in a range of approximately 0...1. |
| 24 | In addition, |
| 25 | it provides the pooled states of the last transformer layer. |
| 26 | The model was created by fine-tuning |
| 27 | [Wav2Vec2-Large-Robust](https://huggingface.co/facebook/wav2vec2-large-robust) |
| 28 | on [MSP-Podcast](https://ecs.utdallas.edu/research/researchlabs/msp-lab/MSP-Podcast.html) (v1.7). |
| 29 | The model was pruned from 24 to 12 transformer layers before fine-tuning. |
| 30 | An [ONNX](https://onnx.ai/) export of the model is available from [doi:10.5281/zenodo.6221127](https://zenodo.org/record/6221127). |
| 31 | Further details are given in the associated [paper](https://arxiv.org/abs/2203.07378) and [tutorial](https://github.com/audeering/w2v2-how-to). |
| 32 | |
| 33 | # Usage |
| 34 | |
| 35 | ```python |
| 36 | import numpy as np |
| 37 | import torch |
| 38 | import torch.nn as nn |
| 39 | from transformers import Wav2Vec2Processor |
| 40 | from transformers.models.wav2vec2.modeling_wav2vec2 import ( |
| 41 | Wav2Vec2Model, |
| 42 | Wav2Vec2PreTrainedModel, |
| 43 | ) |
| 44 | |
| 45 | |
| 46 | class RegressionHead(nn.Module): |
| 47 | r"""Classification head.""" |
| 48 | |
| 49 | def __init__(self, config): |
| 50 | |
| 51 | super().__init__() |
| 52 | |
| 53 | self.dense = nn.Linear(config.hidden_size, config.hidden_size) |
| 54 | self.dropout = nn.Dropout(config.final_dropout) |
| 55 | self.out_proj = nn.Linear(config.hidden_size, config.num_labels) |
| 56 | |
| 57 | def forward(self, features, **kwargs): |
| 58 | |
| 59 | x = features |
| 60 | x = self.dropout(x) |
| 61 | x = self.dense(x) |
| 62 | x = torch.tanh(x) |
| 63 | x = self.dropout(x) |
| 64 | x = self.out_proj(x) |
| 65 | |
| 66 | return x |
| 67 | |
| 68 | |
| 69 | class EmotionModel(Wav2Vec2PreTrainedModel): |
| 70 | r"""Speech emotion classifier.""" |
| 71 | |
| 72 | def __init__(self, config): |
| 73 | |
| 74 | super().__init__(config) |
| 75 | |
| 76 | self.config = config |
| 77 | self.wav2vec2 = Wav2Vec2Model(config) |
| 78 | self.classifier = RegressionHead(config) |
| 79 | self.init_weights() |
| 80 | |
| 81 | def forward( |
| 82 | self, |
| 83 | input_values, |
| 84 | ): |
| 85 | |
| 86 | outputs = self.wav2vec2(input_values) |
| 87 | hidden_states = outputs[0] |
| 88 | hidden_states = torch.mean(hidden_states, dim=1) |
| 89 | logits = self.classifier(hidden_states) |
| 90 | |
| 91 | return hidden_states, logits |
| 92 | |
| 93 | |
| 94 | |
| 95 | # load model from hub |
| 96 | device = 'cpu' |
| 97 | model_name = 'audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim' |
| 98 | processor = Wav2Vec2Processor.from_pretrained(model_name) |
| 99 | model = EmotionModel.from_pretrained(model_name).to(device) |
| 100 | |
| 101 | # dummy signal |
| 102 | sampling_rate = 16000 |
| 103 | signal = np.zeros((1, sampling_rate), dtype=np.float32) |
| 104 | |
| 105 | |
| 106 | def process_func( |
| 107 | x: np.ndarray, |
| 108 | sampling_rate: int, |
| 109 | embeddings: bool = False, |
| 110 | ) -> np.ndarray: |
| 111 | r"""Predict emotions or extract embeddings from raw audio signal.""" |
| 112 | |
| 113 | # run through processor to normalize signal |
| 114 | # always returns a batch, so we just get the first entry |
| 115 | # then we put it on the device |
| 116 | y = processor(x, sampling_rate=sampling_rate) |
| 117 | y = y['input_values'][0] |
| 118 | y = y.reshape(1, -1) |
| 119 | y = torch.from_numpy(y).to(device) |
| 120 | |
| 121 | # run through model |
| 122 | with torch.no_grad(): |
| 123 | y = model(y)[0 if embeddings else 1] |
| 124 | |
| 125 | # convert to numpy |
| 126 | y = y.detach().cpu().numpy() |
| 127 | |
| 128 | return y |
| 129 | |
| 130 | |
| 131 | print(process_func(signal, sampling_rate)) |
| 132 | # Arousal dominance valence |
| 133 | # [[0.5460754 0.6062266 0.40431657]] |
| 134 | |
| 135 | print(process_func(signal, sampling_rate, embeddings=True)) |
| 136 | # Pooled hidden states of last transformer layer |
| 137 | # [[-0.00752167 0.0065819 -0.00746342 ... 0.00663632 0.00848748 |
| 138 | # 0.00599211]] |
| 139 | ``` |