AccentFold aims to enhance Automatic Speech Recognition systems by leveraging the rich diversity of African accents, improving recognition and transcription across multiple dialects.
Utilizing the Afrispeech-200 dataset, which features 120 distinct accents from over 2,463 speakers across 13 African countries, allows for extensive testing and model training.
The implementation of Out Of Distribution (OOD) accents in the testing phase facilitates an evaluation framework for assessing how well AccentFold adapts to previously unencountered accents.
AccentFold's design is grounded in multitask learning techniques, optimizing accent embeddings derived from a pre-trained XLSR model, to improve performance on conversational speech recognition.
#speech-recognition #accent-recognition #machine-learning #african-accents #natural-language-processing
[
Collection
]
[
|
...
]