RobotsMali-AI/afvoices

Git repo for the Bambara African Voices Data Collection and preprocessing # Afvoices — Bambara 600 h Pre‑processing Pipeline This repo collects the **helper utilities** used by our data‑processing team to transform the **626 hours** of raw Bambara speech that we have recorded as part of the African Next Voices (ANV) project into segmented, pre-labeled and share‑ready manifests. Think of it as our audio preprocessing *operations manual*. We share dev-versions of the code, meaning the scripts in this repository are the first implementation and tests of our preprocessing pipeline using Google Cloud Storage FUSE, the code is therefore not optimized for large scale processing but useful if you wish to reproduce the work or have a better understanding of the different components of our pipeline. We have made the raw recordings and their associated metadata and segmentation timestamps publicly available for download, you can download and reconstitute the dataset with the links in manifest/raw-and-meta-600.jsonl. You can also download our SNR evaluation for 612 hours that we processed. --- ## 2 · The Hugging Face Dataset We have processed about 612 hours of those recordings, 159 for which we have corrected the automatically generated transcriptions. We release this segmented version of the dataset on hugging face. A total of 423 hours, for more details please check the dataset card on HF: RobotsMali/afvoices. ### Use this dataset ```python from datasets import load_dataset ds = load_dataset("RobotsMali/afvoices", "human-corrected") # or "model-annotated" / "short" ``` If you want to reconstruct the HF dataset with the original file names you can use the download_from_gcs script and the jsonl files in the manifest folder of this repo to download the segments from GSC. --- ## ASR finetuning experiments We ran Automatic Speech Recognition experiments with a pre-completion subset of this dataset, you can find all the codes and configurations used for those experiments in the dedicated repository: RobotsMali-AI/bambara-asr. All the mode …