evie-8/afrivoices

This repository contains curated subsets of the Digital Umuganda / AfriVoices dataset for the Shona, Lingala, Fulani, and Malagasy languages. The dataset is split into train and test sets for each language.

• Train split: Contains audio clips with corresponding transcriptions.
• Test split: Contains audio clips without transcriptions, as none were available in the original source.

---

Dataset Schema

Each sample in the dataset includes the following fields:

• id: A unique identifier for the audio sample
• audio: The audio clip (in .ogg format)
• audio_language: The language ISO 639-3 code (sna for Shona, lin for Lingala, ful for Fulani, mlg for Malagasy)
• text: The transcription (only present in the train split)
• duration: Duration of the audio clip in seconds
• speaker_id: An anonymized speaker identifier

---

Audio Duration Summary

Language	Code	Total Hours	Train Hours	Test Hours
Shona	sna	574.16 hrs	99.23 hrs	474.92 hrs
Lingala	lin	517.13 hrs	101.52 hrs	415.62 hrs
Fulani	ful	527.45 hrs	124.24 hrs	403.21 hrs
Malagasy	mlg	516.21 hrs	182.51 hrs	333.71 hrs

---

Notes on Data Preparation

This dataset was manually constructed after the official loading script failed due to inconsistent manifest.json structures across language subsets.

• Each language subset had its own format and required custom handling.
• The .tar.xz archives were downloaded and unpacked individually.
• Audio files were matched with metadata to reconstruct valid datasets.
• Some entries were removed because:
  • audio_path was missing in the manifest
  • Audio was too short or corrupted

⚠️ No cleaning or normalization of text has been performed. This is a direct conversion from the original AfriVoices data.

---

Audio Format & Compression

The original .wav audio files from all language subsets totaled well over 750 GB.

To make the dataset more practical for sharing and usage:

• All audio was converted to .ogg format using libvorbis, a high-quality open audio codec
• This reduced the total storage size drastically (e.g., Shona: from ~106 GB to ~14 GB)

If needed, .ogg files can be reconverted back to .wav using tools like ffmpeg.
Original .wav files are available from the AfriVoices dataset.

---

Where to Find the Associated Images

Images that were originally associated with the recordings (via image_path fields) are not included in this dataset. Instead:

• A separate image dataset was created for each language.

• These datasets only include:

  • id: Image identifier
  • image: The image file

• No speaker-level linking is provided (e.g., no speaker_id or guaranteed match to audio).

  You can find the related image datasets here: AfriVoices Image Dataset

---