Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Kunnafonidilaw ka Cadeau: an ASR dataset of present-day Bambara

Yacouba Diarra, Panga Azazia Kamate, Nouhoum Souleymane Coulibaly, Michael Leventhal

TL;DR

This work introduces Kunkado, a 160-hour Bambara ASR dataset derived from Malian radio broadcasts to capture present-day spontaneous speech with code-switching, background noise, and speaker overlap. The authors apply pragmatic transcription normalization and partially human-reviewed corrections, then finetune Parakeet-based Bambara models on a 33.47-hour subset, achieving substantial WER reductions on both a Kunkado test set and a human-evaluated benchmark. The results demonstrate that representativeness and linguistic realism can outweigh sheer data quantity for real-world Bambara ASR, and the authors release both data and models to support broader, code-switching-aware development in low-resource, predominantly oral languages. The work also highlights the challenges of annotating authentic broadcast speech and suggests a community-driven approach to dataset design and evaluation. Overall, Kunkado advances practical Bambara ASR by aligning training data with real-world usage while providing a framework for ongoing, linguistically informed annotation and model improvement.

Abstract

We present Kunkado, a 160-hour Bambara ASR dataset compiled from Malian radio archives to capture present-day spontaneous speech across a wide range of topics. It includes code-switching, disfluencies, background noise, and overlapping speakers that practical ASR systems encounter in real-world use. We finetuned Parakeet-based models on a 33.47-hour human-reviewed subset and apply pragmatic transcript normalization to reduce variability in number formatting, tags, and code-switching annotations. Evaluated on two real-world test sets, finetuning with Kunkado reduces WER from 44.47\% to 37.12\% on one and from 36.07\% to 32.33\% on the other. In human evaluation, the resulting model also outperforms a comparable system with the same architecture trained on 98 hours of cleaner, less realistic speech. We release the data and models to support robust ASR for predominantly oral languages.

Kunnafonidilaw ka Cadeau: an ASR dataset of present-day Bambara

TL;DR

This work introduces Kunkado, a 160-hour Bambara ASR dataset derived from Malian radio broadcasts to capture present-day spontaneous speech with code-switching, background noise, and speaker overlap. The authors apply pragmatic transcription normalization and partially human-reviewed corrections, then finetune Parakeet-based Bambara models on a 33.47-hour subset, achieving substantial WER reductions on both a Kunkado test set and a human-evaluated benchmark. The results demonstrate that representativeness and linguistic realism can outweigh sheer data quantity for real-world Bambara ASR, and the authors release both data and models to support broader, code-switching-aware development in low-resource, predominantly oral languages. The work also highlights the challenges of annotating authentic broadcast speech and suggests a community-driven approach to dataset design and evaluation. Overall, Kunkado advances practical Bambara ASR by aligning training data with real-world usage while providing a framework for ongoing, linguistically informed annotation and model improvement.

Abstract

We present Kunkado, a 160-hour Bambara ASR dataset compiled from Malian radio archives to capture present-day spontaneous speech across a wide range of topics. It includes code-switching, disfluencies, background noise, and overlapping speakers that practical ASR systems encounter in real-world use. We finetuned Parakeet-based models on a 33.47-hour human-reviewed subset and apply pragmatic transcript normalization to reduce variability in number formatting, tags, and code-switching annotations. Evaluated on two real-world test sets, finetuning with Kunkado reduces WER from 44.47\% to 37.12\% on one and from 36.07\% to 32.33\% on the other. In human evaluation, the resulting model also outperforms a comparable system with the same architecture trained on 98 hours of cleaner, less realistic speech. We release the data and models to support robust ASR for predominantly oral languages.
Paper Structure (13 sections, 2 figures, 3 tables)

This paper contains 13 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Characteristics of Kunkado
  3. Audio collection & segmentation:
  4. Noise estimation:
  5. Transcription:
  6. ASR Experiments
  7. Experimental setup:
  8. Training Data:
  9. Training configurations:
  10. Evaluation & Results
  11. Conclusion
  12. Limitations
  13. Acknowledgments

Figures (2)

  • Figure 1: Density Distribution of Signal-to-Noise Ratio values in Kunakdo. This figure includes both subsets
  • Figure 2: WER vs human evaluation. Figure from tall_2025_17672774