Dataset Card for Umini_speech

Dataset Summary

Umini_Speech is a mini_speech dataset developed by the Umbaji team. It consists of over 800 augmented recordings of Ewè words, spoken by various users in different settings. The dataset is designed for speech recognition tasks, particularly for the Yodi model.

Dataset Details

• Size: 800+ recordings
• Language: Ewè
• Audio Format: WAV -Word Classes: gome, djo, miodji, ao, dusi, note, edji, ee
• Corresponding English Labels: down, go, left, no, right, stop, up, yes
• Total Duration: 3.91 hours

Collection Process

• Recordings were collected from 6 users
• Different settings were used for recording to increase diversity
• Audio augmentation techniques were applied to expand the dataset

Usage

Umini_Speech is primarily used to train the Yodi speech recognition model and can also be used for Ewè speech recognition tasks.

Practical Implementation

This is a speech dataset, and users have access to the Yodi model, allowing them to implement code using the Hugging Face library to infer the data. Incorporating code to run inference from our Yodi package is also possible. Keep an eye out for updates as I will push a new version for testing soon.

Additional Information

The paper related to this dataset will be released in September 2024.