Amharic BDU-Speech: 32,901 Paired Audio and Transcriptions

32,901 paired Amharic speech audio files and transcriptions processed from the BDU-speech dataset by Yohannes A. Ejigu. Updated in March 2026, the collection provides mono audio recordings specifically structured for automatic speech recognition research and model training.

Use Cases

• Fine-tuning transformer models like Whisper using the 'audio' and 'sentence' columns
• Developing acoustic models for the Amharic language
• Benchmarking ASR performance on Ethiopian linguistic data

Strengths

• 32,901 training records
• CC BY 4.0 open license
• Processed from the established BDU-speech source

Limitations

• Variable sampling rates across files require normalization
• Limited metadata regarding speaker demographics or recording environments

Provenance

Source

Yohannes A. Ejigu (BDU-speech dataset)

Freshness

Last updated March 2026

Geography

Ethiopia