hadamard-2/alffa-amharic-v2

ALFFA Amharic Speech Corpus (v2)

Read speech corpus for Amharic (አማርኛ) automatic speech recognition. Converted from the original ALFFA project and restructured to match the google/waxalnlp schema for interoperability.

This is a restructured version of hadamard-2/alffa-amharic.

Changes from v1

• utterance_id renamed to id
• transcript renamed to transcription
• speaker_id set to "unknown" — the original ALFFA Kaldi files shipped with utt2spk mapping each utterance to itself, providing no reliable speaker identity information
• language column added, filled with "amh" (ISO 639-2 for Amharic)
• gender column added, filled with "unknown" — not available in the original corpus
• split column dropped — redundant with the HuggingFace dataset split structure
• Schema metadata updated to explicitly declare sampling_rate: 16000 on the audio column

Dataset Structure

{
    'id': 'tr_10000_tr097082',
    'speaker_id': 'unknown',
    'transcription': 'ይህ አማርኛ ጽሑፍ ነው',
    'language': 'amh',
    'gender': 'unknown',
    'audio': {'bytes': ..., 'path': 'tr_10000_tr097082.wav'},  # 16kHz mono
}

Usage

from datasets import load_dataset

dataset = load_dataset("hadamard-2/alffa-amharic-v2")
example = dataset['train'][0]
print(example['transcription'])
audio_array = example['audio']['array']
sampling_rate = example['audio']['sampling_rate']

Splits

• Train: 10,875 utterances (~20 hours)
• Test: 359 utterances (~2 hours)

Citation

@article{tachbelie2014,
    Author = {Martha Tachbelie and Solomon Teferra Abate and Laurent Besacier},
    Journal = {Speech Communication},
    Publisher = {Elsevier},
    Title = {Using different acoustic, lexical and language modeling units for ASR of an under-resourced language - Amharic},
    Volume = {56},
    Year = {2014}
}

License

MIT License (from OpenSLR)