afrospeech-wav2vec-kua

This model is a fine-tuned version of facebook/wav2vec2-base on the crowd-speech-africa, which was a crowd-sourced dataset collected using the afro-speech Space.

Training and evaluation data

The model was trained on a mixed audio data from Oshiwambo (kua).

• Size of training set: 1376
• Size of validation set: 345

Below is a distribution of the dataset (training and valdation)

Evaluation performance

It achieves the following results on the validation set:

• F1: 0.9913480945477086
• Accuracy: 0.9921875

The confusion matrix below helps to give a better look at the model's performance across the digits. Through it, we can see the precision and recall of the model as well as other important insights.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 64
• eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• num_epochs: 150

Training results

Training Loss	Epoch	Validation Accuracy
0.0096	1	0.9843
0.2555	50	0.9843
0.00145	100	0.98177
0.00053	150	0.97770

Framework versions

• Transformers 4.21.3
• Pytorch 1.12.0
• Datasets 1.14.0
• Tokenizers 0.12.1