VERSIONING.md

Key Changes in Version 1.0.1

Jeli-ASR 1.0.1 introduces several updates and enhancements, focused entirely on the transcription side of the dataset. There have been no changes to the audio files since version 1.0.0. Below are the key updates:

1. Symbol Removal:
   All non-vocabulary symbols deemed unnecessary for Automatic Speech Recognition (ASR) were removed, including:
   [ ] ( ) « » ° " < >

2. Punctuation Removal:
   Common punctuation marks were removed to streamline the dataset for ASR use cases. These include:
   : , ; . ? !
   The exception is the hyphen (-), which remains as it is used in both Bambara and French compound words. While this punctuation removal enhances ASR performance, the previous version with full punctuation may still be better suited for other applications. You can still reconstruct the previous version with the archives.

3. Bambara Normalization:
   The transcription were normalized using the Bambara Normalizer, a python package designed to normalize Bambara text for different NLP applications.

4. Optimized Data Format:
   This version introduces .arrow files for efficient data storage and retrieval and compatibility with HuggingFace tools.

These changes enhance the dataset's usability for ASR tasks while providing a cleaner transcription format. Let us know if you have feedback or additional use suggestions for the dataset by opening a discussion or a pull request.

---

Key Changes in Version 1.0.0

1. Name Change

• The Dataset name was changed from jeli-data-manifest to jeli-asr.

2. Mono Channel Conversion

• All stereo audio files have been converted to mono to ensure consistency across the dataset.
  • This step was required only for the jeli-asr-rmai subset as oza-bam-asr was already consistent.

3. Removal of Misaligned Samples

• More than 70% of the data in the previous version contained misaligned samples due to concatenation issues that kind of spread misalignment in the dataset.
• A filtering process was applied using both manual classification and trained classifiers:
  • A subset of the data was manually classified as aligned or misaligned.
  • This subset was used to train classifiers (Logistic Regression and XGBoost) to label the remaining samples. Classifier Performance:
  • Best-performing model: Logistic Regression
    • Accuracy: 0.84
    • F1-score (misaligned - class 0): 0.86
    • F1-score (aligned - class 1): 0.82 Training Details:
  • Balanced training set: Positive samples (aligned) were supplemented using additional aligned samples from Oza's Bambara-ASR dataset.
  • Misaligned samples: No additional samples were needed as they formed a majority.
  • Embedding processing: Manually separated data has been represented as embeddings for training classifiers. The embeddings were obtained by inferring Wav2Vec and BERT, then concatenated for every example and labeled as either aligned or misaligned.

Misaligned samples identified during classification were removed. That subset is currently undergoing further review and may be partially reintegrated in a future version of this dataset.

4. Integration of Oza's Bambara-ASR Dataset

• This version integrates a clean subset from Oza's Bambara-ASR dataset making about 90% of the data.

5. Lowercased Transcriptions

• All transcriptions and translations have been converted to lowercase for consistency.

6. Silent/Empty File Filtering

• Silent or empty audio files with inaudible content were removed.

---