The Makerere AI Lab has built an end-to-end CTC Luganda ASR model using radio data. Having encountered data challenges in working with low resource languages, we take the initiative together with our partners to release the first radio corpus for Luganda.
The corpus of 155 hours is publicly available online under the Creative Commons BY-NC-ND 4.0 license. The dataset release is comprised of the following:
20 hours of human transcribed radio speech. The audio is 16kHZ, mono channel and with 16 bit rate.
Two CSV files for the 20-hour human transcribed dataset - cleaned.csv contains cleaned transcripts and uncleaned.csv contains uncleaned transcripts. The uncleaned transcripts contain extra speech details included in tags like [laughter] for laughter, and [um] for filler pauses, which speaker is talking, where each speaker is assigned an identifier A or B.
The transcription guide used to transcribe the radio dataset.
A multi-speaker untranscribed dataset of 6 hours of radio data. 1.4 hours of women voices and 4.6 hours of men voices. Each audio is a ten-seconds clip with a single speaker.
135 hours of multi-speaker untranscribed radio data.