osinkolu/fongbe-hausa-asr

# Fongbe ASR Dataset Creator A pipeline for building a unified Automatic Speech Recognition (ASR) dataset for **Fongbe**, a low-resource language spoken in Benin. The notebook combines two open data sources into a single, train/validation/test dataset and publishes it to the Hugging Face Hub. ## Motivation Fongbe is significantly underrepresented in existing speech technology. This project consolidates the available open Fongbe speech data into one standardized dataset to support ASR model training and evaluation. ## Data Sources | Source | Origin | Samples | Split | |--------|--------|---------|-------| | Zenodo Fongbe Speech Dataset | Crowdsourced recordings with metadata (speaker, gender, age, dialect) | 3,179 | validation | | ALFFA Fongbe (train) | Kaldi-formatted read speech corpus | 8,234 | train | | ALFFA Fongbe (test) | Kaldi-formatted read speech corpus | 2,168 | test | ## Dataset Statistics | Split | Samples | Duration | |-------|---------|----------| | Train | 8,234 | 5.73 h | | Validation | 3,179 | 5.11 h | | Test | 2,168 | 1.45 h | | **Total** | **13,581** | **12.30 h** | ## Published Dataset The merged dataset is available on the Hugging Face Hub: **`Professor/fongbe-speech-zenodo`** Schema (all splits): | Column | Type | Description | |--------|------|-------------| | `audio` | Audio (16 kHz) | Waveform array with sampling rate | | `text` | string | Transcription | | `speaker ID` | string | Speaker identifier | | `audio filename` | string | Original WAV filename | ## Notebook Workflow The pipeline in `Fon_ASR_Dataset_creator.ipynb` follows these steps: 1. **Download Zenodo data** — fetches and unzips the Fongbe Speech Dataset archive 2. **Load metadata** — reads the CSV with speaker demographics and file mappings 3. **Compare with existing HF dataset** — loads `beethogedeon/fongbe-speech` and compares transcripts against the Zenodo data 4. **Investigate differences** — character-level and substring analysis reveals the HF dataset contai …