RobotsMali/afvoices

📘 African Next Voices – Bambara (AfVoices)

The AfVoices dataset is the largest open corpus of spontaneous Bambara speech at its release in late 2025. It contains 423 hours of segmented audio and 612 hours of original raw recordings collected across southern Mali. Speech was recorded in natural, conversational settings and annotated using a semi-automated transcription pipeline combining ASR pre-labels and human corrections. We release all the data processing code on GitHub.

🔎 Quick Facts

Load the Dataset

Choose a configuration explicitly because the subsets have different supervision levels:

from datasets import load_dataset

reviewed = load_dataset(
    "RobotsMali/afvoices",
    "human-corrected",
    split="train",
)
automatic = load_dataset(
    "RobotsMali/afvoices",
    "model-annotated",
    split="train",
)

Use text as the validated reference only in human-corrected. The label-v1 and label-v2 fields are model-generated hypotheses and must not be presented as human ground truth. The ASR results reported in the accompanying paper and v2 model cards used an earlier pre-completion subset; its exact manifests are archived in the `bambara-asr` repository.

Motivation

The African Next Voices (ANV) project is a multi-country effort aiming to gather over 9,000 hours of speech across 18 African languages. Its goal is to build high-quality datasets that empower local communities, support inclusive AI research, and provide strong foundations for ASR in underrepresented languages.

As part of this initiative, RobotsMali led the Bambara data collection for Mali. This dataset reflects RobotsMali’s broader mission to advance AI and NLP research malian languages, with a long-term focus on improving education, access, and technology across Mali and the wider Manding linguistic region.

🎙️ Characteristics of the Dataset

Data Collection

• —Speech was collected through trained facilitators who guided participants, ensured audio quality, and encouraged natural, topic-focused conversations.
• —All recordings are spontaneous speech, not read text.
• —A custom Flutter mobile app (open-source) was used to simplify the process and reduce training time.
• —Geographic focus: Southern Mali, to limit extreme accent variation and build a clean baseline corpus.

Segmentation and Preprocessing

• —Raw audio was segmented using Silero VAD, retaining ~70% of the original duration.
• —Segments range from 240 ms to 30 s.
• —Voice activity detection helped remove long silences and improve data usability.

Transcriptions

• —Pre-transcribed using the ASR model soloni-114m-tdt-ctc-v0.
• —Human annotators corrected the transcripts.
• —A second model (soloni-114m-tdt-ctc-v2) was trained using the corrected transcripts and used to regenerate improved labels.
• —Two automatic transcription variants exist for each sample: v1 (from soloni-v0) and v2 (from soloni-v2).

Acoustic Event Tags

The following tags appear in transcriptions:

📂 Subsets

1. Human-corrected (159 h, 260k samples)

• —Fully reviewed and corrected by annotators.
• —Only subset with a definitive text field containing the validated transcription.

2. Model-annotated (212 h, 355k samples)

• —Includes automatic labels: v1 (soloni-v0) and v2 (soloni-v2).
• —No human review.

3. Short subset (52 h, 259k samples)

• —Segments <1 second (formulaic expressions, discourse markers).
• —Excluded from human annotation for optimization purposes.
• —Automatically labeled (v1 & v2).

⚠️ Limitations

• —Clean dataset vs real-world noise: Over 70% of recordings can be categorized as relatively clean speech. Models trained solely on this dataset may underperform in noisy street or radio environments typical in Mali. See this report if you are interested in learning more about the strengths and weaknesses of RobotsMali's ASR models.

• —Reduced code-switching: French terms were often replaced by [cs] or normalized into Bambara phonology. This improves model stability but reduces realism for natural bilingual speech.

• —Geographic homogeneity: Focused on the southern region to control accent variability. Broader dialectal coverage might require additional data.

• —Simplified linguistic conditions: Overlaps, multi-speaker settings, and conversational chaos are minimized—again improving training stability at the cost of deployment realism.

📑 Citation

@misc{diarra2025dealinghardfactslowresource,
      title={Dealing with the Hard Facts of Low-Resource African NLP},
      author={Yacouba Diarra and Nouhoum Souleymane Coulibaly and Panga Azazia Kamaté and Madani Amadou Tall and Emmanuel Élisé Koné and Aymane Dembélé and Michael Leventhal},
      year={2025},
      eprint={2511.18557},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.18557},
}

You may want to download the original 612 hours dataset with its associated metadata for research purposes or to create a derivative. You will find the codes and manifest files to download those files from Google Cloud Storage in this repository: RobotsMali-AI/afvoices. Do not hesitate to open an issue for Help or suggestions 🤗