This repository contains the 2 hours speech recognition dataset recorded in Bemba language. The native language spoken by the majority of people in the north part of Zambia. The dataset is split into 1hr dev
and 1hr test
datasets as per directive by Prof. Laurent Besacier.
*.wav
audio files - produced by the Lig-Aikuma application*.json
metadata files - produced by the Lig-Aikuma application associated with every audio file.*linker.text
files - produced by Lig-Aikuma application after recording session.Session.text
files - consists raw texts used for elicitation.
- [Bemba_speech_dataset] --> ['dev', 'train']
Several challenges were encoutered during the dataset building process:
- The scarcity of the Bemba literature on the internet. But was fortunate to find an online Bible in Bemba language.
- Difficult to read and pronounce some of the words in Bemba. Surely, this could have negatively affected the quality of recordings.
- The Lig-Aikuma application was crushing during recording sessions. This slowed the process of recording and frustrating at times.
- Speech recognition course was very interesting especially that it was coupled with a project that gave me the practical aspect of gathering data and coming up with speech dataset for ASR. Huge shout outs to amazing lecturers for sharing their knowledge.
- Having noticed there exist no speech dataset for any Zambia language. I wish to take up a challenge/project and build the speech datasets for some of the local language in my country, Zambia, especially after AMMI. This would allow me and other researcher interested in speech recognition to carry on research. Also this has an impact on preservation of some local languages.
-- Signed
- Claytone Sikasote | csikasote@aimsammi.org