Yoruba Speech Dataset for Low-Resource Speech Tasks

This repo contains over 3 hours of recorded yoruba text from 1 native yoruba male speaker.

Downloading

Go to your terminal and enter;

git clone https://github.com/ogunlao/yoruba_speech_project.git

This adds a folder called "yoruba_speech_project" which contain the files to your local directory.

Statistics

• 1079 recordings of a maximum of 20 word length per recording
• 3.06 total recording hours
• 22 folders each containing 49 recorded audio with their respective metadata.

Recordings are divided into folders for easy access.

Data Collection

Text dataset used for recording was collected from Niger Volta LTI. It had a smooth narrative which made the text interesting to read.

Recording was done using the Lig-Aikuma Android app. It is an easy-to-use app with good interface for recording and elicitation. It offers 6 modes of usage;

• Recording
• Respeaking
• Translating
• Elicitation
• Check
• Share

The elicitation mode was used in data collection where a displayed text is read carefully by the speaker during recording.

Data Preprocessing

Preprocessing involved;

• validating data for errors and removing corrupt files
• merging folders
• splitting data into train, val and test samples

Two main dataset directory with subdirectories;

• recordings contain the unprocessed recorded files and metadata
• data/records contain split data

Raw speech dataset split into;

• 1 hour of speech data (split into train and val)
• 1 hour of speech as test set
• Additional extra 1 hour of speech as extra data
• each contained in data/records/train, data/records/val, data/records/extra respectively

All preprocessing can be accessed through the notebook yoruba_speech_preprocessing and script yor_processor.py

Note that a python installation is required with some dependencies;

• numpy
• wave

Application

The dataset can be used majorly for low-resource speech model experiments or for cross-lingual ASR. Also, a pretrained model can be fine-tuned on this dataset to learn phonetic representations for yoruba language.

Problems Encountered

• Difficulty in avoiding noise from environment when recording. Some recordings had to be performed multiple times to get a clearer audio
• Few audio files were corrupted after recording. I filtered them out during preprocessing. Be aware of this when using the raw (unprocessed) audio files.
• Some words in text were difficult to pronounce without proper accents.
• Recording speech takes time and can become uninteresting to perform quickly.

Contributing

If you will like to contribute to this project by recording more audio files, I have text files under slit_text folder where you can easily read short sentences. Upload the text files to the lig-aikuma android app and start recording. Currently at yor_split_21 for this project. You can make a pull request and I will be happy to add you to the project.

Original Author

Sewade Olaolu Ogun

License

MIT