edaiofficial/okwugbe

Automatic Speech Recognition for African Language ## Okwugbe Automatic Speech Recognition Library for (low-resource) African Languages ## Motivation Our aim is to foster ASR for African languages by making the whole process--from dataset gathering and preprocessing to training--as easy as possible. This library follows our work Okwugbé on ASR for Fon and Igbo. Based on the architecture of the network described in our paper, it aims at easing the training process of ASR for other languages. The primary targets are African languages, but it supports other languages as well ## Usage ```pip install okwugbe``` ```python #Import the trainer instance from train_eval import Train_Okwugbe train_path = '/path/to/training_file.csv' test_path = '/path/to/testing_file.csv' characters_set = '/path/to/character_set.txt' """ /path/to/training_file.csv and /path/to/testing_file.csv are meant to be csv files with two columns: the first one containing the full paths to audio wav files the second one containing the textual transcription of audio contents """ #Initialize the trainer instance train = Train_Okwugbe(train_path, test_path, characters_set) #Start the training train.run() ``` ### Parameters Here are the parameters for the package, as well as their default values. The default values have been chosen so that you only have to make minimal changes to get a good ASR model going. | Parameter | Description | default | | --- | --- | --- | | `use_common_voice` | Whether or not to use common voice | False | | `lang` | language to use from Common Voice. Must be specified if `use_common_voice` is set to True. | None | | `rnn_dim` | RNN Dimension & Hidden Size | 512 | | `num_layers` | Number of Layers | 1 | | `n_cnn` | Number of CNN components | 5 | | `n_rnn` | Number of RNN components | 3 | | `n_feats` | Number of features for the ResCNN | 128 | | `in_channels` | Number of input channels of the ResCNN | 1 | | `out_channels` | Number of output channels of the ResCNN | 32 | | `ke …