RobotsMali/afvoices-notag

AfVoices Top-20 Speakers without Tags

RobotsMali/afvoices-notag is a small experimental TTS-oriented selection derived from RobotsMali/afvoices, the African Next Voices Bambara speech corpus. It contains the 20 participants with the highest utterance counts and excludes transcripts containing semantic/acoustic annotation tags.

This is the dataset used for RobotsMali's first Bambara VITS experiments. It is not a high-quality studio TTS corpus: the source is spontaneous speech collected and annotated primarily for ASR. The subset was judged sufficient to test training and text-representation hypotheses, not to build a production voice.

Quick facts

Item	Value
Configuration	top20-speakers
Train examples	21,253
Test examples	1,128
Selected speakers	20
Language	Bambara (bm)
Speech style	Spontaneous
Audio storage	Public URLs in the audio string column

The small Parquet files contain metadata and audio URLs, not embedded waveforms. Network access is therefore required to retrieve audio, and long-term use depends on those URLs remaining available.

Load the dataset

from datasets import Audio, load_dataset

dataset = load_dataset(
    "RobotsMali/afvoices-notag",
    "top20-speakers",
)

# Decode and resample URL-backed audio when it is accessed.
dataset = dataset.cast_column("audio", Audio(sampling_rate=22_050))
sample = dataset["train"][0]
print(sample["text"], sample["participantId"])

Fields

• text (string): Bambara transcript after removal of examples containing semantic/acoustic tags
• audio (string): URL of the audio segment; cast it to datasets.Audio for decoding
• participantId (string): source participant identifier; treat it as a pseudonymous identifier, not a name
• split (string): source split label, duplicating the Dataset split organization

Creation and relationship to AfVoices

The parent AfVoices release contains spontaneous speech collected in southern Mali through the African Next Voices project. Its transcription workflow combined automatic pre-labeling and human correction. This derivative selects the top 20 speakers by utterance count and removes examples whose transcripts include the semantic/acoustic tags used in AfVoices, such as markers for noise, code-switching, or inaudible speech.

Removing tagged transcripts does not guarantee clean audio: untagged background noise, recording variation, disfluencies, and transcription errors can remain. Refer to the full AfVoices card and processing repository for collection context.

Intended use and limitations

The dataset is intended for exploratory TTS/ASR work, data-pipeline testing, and reproduction of the bam-vits experiments. It should not be described as balanced, representative, studio quality, or sufficient for production TTS.

• Selecting by utterance count overrepresents frequent contributors and excludes the broader AfVoices speaker distribution.
• The train/test split is not documented as a speaker-disjoint evaluation.
• Spontaneous ASR recordings have inconsistent prosody, environment, microphones, and transcript fidelity for TTS.
• Excluding tagged transcripts introduces selection bias and does not prove absence of acoustic events.
• The dataset may contain regional, demographic, topic, and recording-condition imbalance inherited from AfVoices.
• The public participantId field and recognizable voices require privacy-aware use. Do not attempt to identify speakers or build impersonation systems.

Users should listen to samples, audit text/audio alignment, check consent and licensing for their use case, and document any additional filtering.

Models trained on this dataset

This subset trained the base bam-vits and bam-vits-pseudo-ipa experimental lines. Both are substantially undertrained and have no formal perceptual evaluation.

Citation

Please cite the parent African Next Voices work and identify this derivative by repository ID.

@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}
}

Questions are welcome on the dataset Community tab or in the AfVoices repository.