README.md
| 1 | --- |
| 2 | datasets: |
| 3 | - agender |
| 4 | - mozillacommonvoice |
| 5 | - timit |
| 6 | - voxceleb2 |
| 7 | inference: true |
| 8 | tags: |
| 9 | - speech |
| 10 | - audio |
| 11 | - wav2vec2 |
| 12 | - audio-classification |
| 13 | - age-recognition |
| 14 | - gender-recognition |
| 15 | license: cc-by-nc-sa-4.0 |
| 16 | base_model: |
| 17 | - facebook/wav2vec2-large-robust |
| 18 | --- |
| 19 | |
| 20 | # Model for Age and Gender Recognition based on Wav2vec 2.0 (24 layers) |
| 21 | |
| 22 | The model expects a raw audio signal as input and outputs predictions |
| 23 | for age in a range of approximately 0...1 (0...100 years) |
| 24 | and gender expressing the probababilty for being child, female, or male. |
| 25 | In addition, it also 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 [aGender](https://paperswithcode.com/dataset/agender), |
| 29 | [Mozilla Common Voice](https://commonvoice.mozilla.org/), |
| 30 | [Timit](https://catalog.ldc.upenn.edu/LDC93s1) and |
| 31 | [Voxceleb 2](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/vox2.html). |
| 32 | For this version of the model we trained all 24 transformer layers. |
| 33 | An [ONNX](https://onnx.ai/) export of the model is available from |
| 34 | [doi:10.5281/zenodo.7761387](https://doi.org/10.5281/zenodo.7761387). |
| 35 | Further details are given in the associated [paper](https://arxiv.org/abs/2306.16962) |
| 36 | and [tutorial](https://github.com/audeering/w2v2-age-gender-how-to). |
| 37 | |
| 38 | # Usage |
| 39 | |
| 40 | ```python |
| 41 | import numpy as np |
| 42 | import torch |
| 43 | import torch.nn as nn |
| 44 | from transformers import Wav2Vec2Processor |
| 45 | from transformers.models.wav2vec2.modeling_wav2vec2 import ( |
| 46 | Wav2Vec2Model, |
| 47 | Wav2Vec2PreTrainedModel, |
| 48 | ) |
| 49 | |
| 50 | |
| 51 | class ModelHead(nn.Module): |
| 52 | r"""Classification head.""" |
| 53 | |
| 54 | def __init__(self, config, num_labels): |
| 55 | |
| 56 | super().__init__() |
| 57 | |
| 58 | self.dense = nn.Linear(config.hidden_size, config.hidden_size) |
| 59 | self.dropout = nn.Dropout(config.final_dropout) |
| 60 | self.out_proj = nn.Linear(config.hidden_size, num_labels) |
| 61 | |
| 62 | def forward(self, features, **kwargs): |
| 63 | |
| 64 | x = features |
| 65 | x = self.dropout(x) |
| 66 | x = self.dense(x) |
| 67 | x = torch.tanh(x) |
| 68 | x = self.dropout(x) |
| 69 | x = self.out_proj(x) |
| 70 | |
| 71 | return x |
| 72 | |
| 73 | |
| 74 | class AgeGenderModel(Wav2Vec2PreTrainedModel): |
| 75 | r"""Speech emotion classifier.""" |
| 76 | |
| 77 | def __init__(self, config): |
| 78 | |
| 79 | super().__init__(config) |
| 80 | |
| 81 | self.config = config |
| 82 | self.wav2vec2 = Wav2Vec2Model(config) |
| 83 | self.age = ModelHead(config, 1) |
| 84 | self.gender = ModelHead(config, 3) |
| 85 | self.init_weights() |
| 86 | |
| 87 | def forward( |
| 88 | self, |
| 89 | input_values, |
| 90 | ): |
| 91 | |
| 92 | outputs = self.wav2vec2(input_values) |
| 93 | hidden_states = outputs[0] |
| 94 | hidden_states = torch.mean(hidden_states, dim=1) |
| 95 | logits_age = self.age(hidden_states) |
| 96 | logits_gender = torch.softmax(self.gender(hidden_states), dim=1) |
| 97 | |
| 98 | return hidden_states, logits_age, logits_gender |
| 99 | |
| 100 | |
| 101 | |
| 102 | # load model from hub |
| 103 | device = 'cpu' |
| 104 | model_name = 'audeering/wav2vec2-large-robust-24-ft-age-gender' |
| 105 | processor = Wav2Vec2Processor.from_pretrained(model_name) |
| 106 | model = AgeGenderModel.from_pretrained(model_name) |
| 107 | |
| 108 | # dummy signal |
| 109 | sampling_rate = 16000 |
| 110 | signal = np.zeros((1, sampling_rate), dtype=np.float32) |
| 111 | |
| 112 | |
| 113 | def process_func( |
| 114 | x: np.ndarray, |
| 115 | sampling_rate: int, |
| 116 | embeddings: bool = False, |
| 117 | ) -> np.ndarray: |
| 118 | r"""Predict age and gender or extract embeddings from raw audio signal.""" |
| 119 | |
| 120 | # run through processor to normalize signal |
| 121 | # always returns a batch, so we just get the first entry |
| 122 | # then we put it on the device |
| 123 | y = processor(x, sampling_rate=sampling_rate) |
| 124 | y = y['input_values'][0] |
| 125 | y = y.reshape(1, -1) |
| 126 | y = torch.from_numpy(y).to(device) |
| 127 | |
| 128 | # run through model |
| 129 | with torch.no_grad(): |
| 130 | y = model(y) |
| 131 | if embeddings: |
| 132 | y = y[0] |
| 133 | else: |
| 134 | y = torch.hstack([y[1], y[2]]) |
| 135 | |
| 136 | # convert to numpy |
| 137 | y = y.detach().cpu().numpy() |
| 138 | |
| 139 | return y |
| 140 | |
| 141 | |
| 142 | print(process_func(signal, sampling_rate)) |
| 143 | # Age female male child |
| 144 | # [[ 0.33793038 0.2715511 0.2275236 0.5009253 ]] |
| 145 | |
| 146 | print(process_func(signal, sampling_rate, embeddings=True)) |
| 147 | # Pooled hidden states of last transformer layer |
| 148 | # [[ 0.024444 0.0508722 0.04930823 ... 0.07247854 -0.0697901 |
| 149 | # -0.0170537 ]] |
| 150 | ``` |