Hugging Face's logo

Datasets:
Tamazight-NLP
/
TOSD

Tasks:
Automatic Speech Recognition
Text-to-Speech
Modalities:
Audio
Text
Formats:
parquet
Languages:
Standard Moroccan Tamazight
ber
Size:
1K - 10K
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
TOSD / README.md
ayymen's picture
ayymen
Update README.md
8bbbc8e verified
preview code
|
raw
history blame contribute delete
4.4 kB
metadata 
dataset_info:
  features:
    - name: text
      dtype: string
    - name: subset
      dtype: string
    - name: audio
      dtype: audio
    - name: duration
      dtype: float64
  splits:
    - name: train
      num_bytes: 395694519.261
      num_examples: 1801
  download_size: 587038815
  dataset_size: 395694519.261
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
language:
  - zgh
  - ber
task_categories:
  - automatic-speech-recognition
  - text-to-speech
pretty_name: Tamazight Open Speech Dataset
license: apache-2.0

Dataset Card for Tamazight Open Speech Dataset

This dataset provides a parsed, formatted, and ready-to-use Amazigh Voice Dataset. It contains voice recordings and corresponding text transcripts in Standard Moroccan Amazigh (ⵜⴰⵎⴰⵣⵉⵖⵜ ⵜⴰⵏⴰⵡⴰⵢⵜ ⵜⴰⵎⵓⵔⴰⴽⵓⵛⵜ) intended for training Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) models.

This specific repository is published by a collaborator. You may visit the raw dataset repository which has additional dataset that hasn't been uploaded here yet: Amazigh-Speech-Dataset.

Dataset Details

Dataset Description

Dataset Sources

Uses

Direct Use

  • Training or fine-tuning Speech-to-Text (STT / ASR) models.
  • Training or fine-tuning Text-to-Speech (TTS) models.
  • Linguistic research regarding Amazigh phonetics and speech.

Out-of-Scope Use

This data should not be used to generate malicious voice clones or deepfakes intended for impersonation, fraud, or harassment.

Dataset Structure

Unlike the raw dataset which uses separate .wav and .txt files in .zip archives, this dataset has been parsed into a structured format for immediate use with the Hugging Face datasets library.

It contains 1,801 examples with the following fields:

  • audio: The audio data feature, containing the decoded audio array and sampling rate.
  • text: The string transcript of the audio in the Tifinagh script.
  • duration: The length of the audio clip in seconds (float64).
  • subset: The dataset has two different subsets recorded using different microphones. While subset_1 has mono audio, subset_2 is stereo.

Dataset Creation

Curation Rationale

Amazigh is a low-resource language in AI. This dataset was created to contribute high-quality, openly licensed voice data to help the open-source community build better voice technologies for the Amazigh-speaking community.

Source Data

Data Collection and Processing

The audio was recorded by a fluent speaker reading from pre-selected Standard Moroccan Amazigh texts. This specific repository hosts the processed and structured version of those original recordings.

Who are the source data producers?

The audio was originally recorded by Abdelhaque Id Ali, a speaker of Southern Moroccan Amazigh.

Personal and Sensitive Information

The dataset contains the voice recordings of the creator. No other personally identifiable information (PII) is included in the audio or text.

Bias, Risks, and Limitations

This dataset represents the voice, accent, and pronunciation of a single speaker using Standard Moroccan Amazigh. It may not fully capture the phonetic diversity of other regional Amazigh varieties. Models trained solely on this data may struggle with accents or dialects not represented here.

Dataset Card Authors

Dataset Card Contact