ibso99/Amharic-ASR-

Am Amharic speech to text or ASR model trained on 205.17 hours of data # Amharic-ASR- # Amharic ASR Dataset EDA This project performs exploratory data analysis on a merged Amharic ASR training dataset created from Waxal and Librispeech-style Amharic speech data. The goal was to validate the dataset before fine-tuning a Whisper-based speech recognition model. ## Dataset - Total samples: 39,022 - Split analyzed: Train - Sources: Waxal + Librispeech-style Amharic data - Target model: Whisper - Target sample rate: 16 kHz ## Key Findings - 99.9% of audio files are within the 0.5–30 second target range - Median duration: 17.38 seconds - Mean duration: 17.64 seconds - 100% of inspected files use the target 16 kHz sample rate - 0 empty transcripts - 98.5% pure Amharic transcripts - 566 unique speakers - 515 duplicate transcripts require leakage checks - 20 files are longer than 30 seconds - 1,000 samples have missing language, gender, and speaker metadata ## Conclusion The dataset is suitable for Whisper fine-tuning after minor preprocessing. Recommended steps include duration filtering, mixed-script transcript review, duplicate transcript inspection, speaker-aware splitting, and final audio quality validation.