electroducer/acholi4kaldi

A Kaldi recipe for Acholi data # Acholi for Kaldi ## A complete Kaldi recipe for training an Acholi ASR system ### About This recipe is based on existing Kaldi recipes and was produced as part of a dissertation project at the University of Edinburgh. The scripts will make baseline ASR systems for Acholi, a language spoken in northern Uganda and South Sudan. A series of two-dimensional experiments are also carried out to compare phoneme-based vs. grapheme-based acoustic models as well as various bottleneck features trained on supplementary datasets from multiple languages. ### Required data The scripts are designed for proprietary datasets and will not run as is without them. The Acholi and supplementary Luganda radio datasets were obtained from a third party and cannot be acquired without express permission. The supplementary Swahili dataset is from the GlobalPhone corpus, which can be purchased. The scripts will expect the source data to be found in a parallel directory called `data`; however, this path can be changed by modifying the `DATA_PATH` variable in the `path.sh` script. ### Required links Similar to other Kaldi recipes, local symbolic links pointing to the `steps` and `utils` directories from the WSJ Kaldi recipes are required. ### Running the scripts For all `run*.sh` scripts, it is recommended that they be run line-by-line in order to ensure that they work as expected and to break up the training and decoding process, which may take a considerable amount of time. - `run_{1,2,3}_gmm_{acholi,swahili,luganda}.sh`: These scripts will train and decode a series of baseline HMM-GMM systems for for Acholi, Swahili, and Luganda, respectively. The scripts are almost identical; however, several fundamental differences in the data among languages, they have been separated into multiple scripts rather than looping over each language in a single script. - `run_4_advanced.sh`: This script will first train and decode a baseline HMM-DNN system, then it will train various bottleneck DNNs, e …