dmatekenya/Chichewa-Speech2Text

Automated Speech Recognition for Chichewa. # Automated Speech Recognition(ASR) for Chichewa This repository shares results, code and activities for work whose goal is to contribute to improvements in ASR or speech to text technologies (e.g., NLP models) for Chichewa which happens to be my native language and also the major language spoken in Malawi. In Zambia and Mozambique, Chichewa is also known as *Chinyanja*, *Nyanja* or *Chewa*. At the time of writing of this document, over 21 million people speak the language. The project has two main goals as follows: 1. **Generate speech and other NLP datasets.** Contribute to generation of open source speech datasets for the language. Since quality of results from speech systems is inevitably connected to availability and quality of other NLP datasets in the language, the project is also interested in contributing to other NLP datasets such as those for translations. 2. **Enhance ASR models.** Since Chichewa is a low resource language, the quality of speech to texts results from platforms such as Google is poor. As such, the goal here is to fine-tune these large language models with properly curated data to improve performance of the models. Again, this applies to other downstream NLP tasks such as *machine translation*, *text summarization* and *question answering*. Furthermore, improving accessibility and usage of these models among locals in the target countries is another important objective. The repository contains the following folders: - **docs/** > This folder contains documents pertaining to the project and the repo itsself. - **data/** > The datasets in this folder will not be uploaded to the repo as most of them are too large. However, finalized datasets will be shared through platforms such as Zenodo. - **src/** > All the Python and other code used is placed here. For example, this folder contains source code for the speech data collection app: LIG-Akuma - **models/** > Any models available after tuning will be here - **notebooks/** > All …