Dvoice-v2-cleaned / README.md
BounharAbdelaziz's picture
Update README.md
3856fa4 verified
metadata
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: transcription_darija_ar
      dtype: string
    - name: transcription_darija_ltn
      dtype: string
    - name: intent
      dtype: string
    - name: sentiment
      dtype: string
    - name: gender_label
      dtype: string
    - name: lead_time
      dtype: float64
    - name: sample_rate
      dtype: int64
    - name: split
      dtype: string
    - name: duration
      dtype: float64
  splits:
    - name: train
      num_bytes: 16650545
      num_examples: 120
  download_size: 15088227
  dataset_size: 16650545
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Overview

This dataset is a cleaned version of the Dvoice-v2 test set, curated and enhanced for easier use in speech and language processing tasks. The original dataset contains many errors, with audios not matching their transcripts.

This dataset has been manually annotated by Abdelaziz Bounhar, with the following additional labels:

  • Arabizi transcription: Transcription in latin scripts.
  • Gender: Classification of speaker's gender.
  • Sentiment: Sentiment label of the speaker's statement.
  • Intent: The intent behind the speaker's statement, helping to better understand the context.

Code switching (in the arabic transcription) was added when necessary.

This dataset is ready for use in tasks such as speech emotion recognition, speaker profiling, and intent classification, making it a versatile resource evaluating machine learning models.