README.md

---
dataset_info:
- config_name: dioula
  features:
  - name: audio
    dtype:
      audio:
        decode: false
  - name: text
    dtype: string
  - name: language
    dtype: string
  - name: duration
    dtype: string
  - name: chunk_id
    dtype: string
  - name: record_id
    dtype: string
  - name: content_type
    dtype: string
  splits:
  - name: train
    num_bytes: 86240632
    num_examples: 966
  - name: validation
    num_bytes: 9552535
    num_examples: 107
  download_size: 95739119
  dataset_size: 95793167
- config_name: fulfulde
  features:
  - name: audio
    dtype:
      audio:
        decode: false
  - name: text
    dtype: string
  - name: language
    dtype: string
  - name: duration
    dtype: string
  - name: chunk_id
    dtype: string
  - name: record_id
    dtype: string
  - name: content_type
    dtype: string
  splits:
  - name: train
    num_bytes: 280046437
    num_examples: 2034
  - name: validation
    num_bytes: 31116202
    num_examples: 226
  download_size: 311038493
  dataset_size: 311162639
- config_name: moore
  features:
  - name: audio
    dtype:
      audio:
        decode: false
  - name: text
    dtype: string
  - name: language
    dtype: string
  - name: duration
    dtype: string
  - name: chunk_id
    dtype: string
  - name: record_id
    dtype: string
  - name: content_type
    dtype: string
  splits:
  - name: train
    num_bytes: 425869746
    num_examples: 5171
  - name: validation
    num_bytes: 47355803
    num_examples: 575
  download_size: 472982755
  dataset_size: 473225549
configs:
- config_name: dioula
  data_files:
  - split: train
    path: dioula/train-*
  - split: validation
    path: dioula/validation-*
- config_name: fulfulde
  data_files:
  - split: train
    path: fulfulde/train-*
  - split: validation
    path: fulfulde/validation-*
- config_name: moore
  data_files:
  - split: train
    path: moore/train-*
  - split: validation
    path: moore/validation-*
language:
- mos
- dyu
- ful
task_categories:
- automatic-speech-recognition
license: cc-by-4.0
---

Faso Speech is a speech dataset built for Burkina Faso languages, starting with
paired audio and text from Moore Burkina resources.

The current target languages are:

- Moore
- Dioula
- Fulfulde

French rows are preserved when they appear in the processed source material,
mostly as translation or supporting content.

## Dataset Summary

The dataset contains short audio chunks paired with transcripts. Source pages
were archived separately, then processed into timed candidate chunks for ASR
training and review.

The first source family is Moore Burkina. Many records embed or link IPS
app-builder pages hosted on `media.ipsapps.org`; those pages expose timing
arrays, text blocks, audio URLs, and next-page links used during extraction.

## Current Processed Totals

These totals describe the processed dataset metadata used for this release.

| Language | Rows | Duration | Rows shorter than 0.72s |
| --- | ---: | ---: | ---: |
| moore | 5,746 | 05:41:41.870 | 88 |
| fulfulde | 2,260 | 02:41:47.140 | 103 |
| dioula | 1,073 | 00:49:45.910 | 26 |
| french | 212 | 00:12:23.400 | 4 |

## Configurations

Use one configuration per primary language:

- `moore`
- `dioula`
- `fulfulde`

Each configuration has deterministic `train` and `validation` splits. The
split ratio is 90% train and 10% validation.

## Data Fields

The dataset uses a compact ASR-oriented schema:

| Column | Description |
| --- | --- |
| `audio` | Embedded Hugging Face `Audio(decode=False)` value with `bytes` and `path` |
| `text` | Transcript |
| `language` | Language label |
| `duration` | Chunk duration from metadata |
| `chunk_id` | Stable chunk identifier |
| `record_id` | Source record identifier |
| `content_type` | Source content type |

## Usage

```python
from datasets import Audio, load_dataset

ds = load_dataset("madoss/faso-speech", "moore")
ds = ds.cast_column("audio", Audio(decode=False))

print(ds["train"][0]["text"])
print(ds["train"][0]["audio"])
```

## Audio Preparation

The dataset is distributed as embedded Hugging Face `Audio(decode=False)`
examples. When segment metadata was available during preparation, only detected
speech ranges were embedded, with small padding around speech to avoid clipping
words. Rows without segment metadata keep their original chunk audio.

Preparation used the following default padding and merge behavior:

| Option | Default |
| --- | ---: |
| `--segment-start-padding` | 0.15 |
| `--segment-end-padding` | 0.25 |
| `--music-start-padding` | 0.05 |
| `--music-end-padding` | 0.05 |
| `--max-intra-segment-gap` | 0.50 |
| `--max-music-gap` | 0.05 |

## Source Data And Provenance

The source archive preserves raw source artifacts separately from processed
training examples, including source URLs, app-builder HTML, and optional wrapper
HTML, audio URLs, downloaded audio, extracted text blocks, timing metadata,
language labels, content type, and metadata.

This separation keeps the processed dataset traceable to the original source
material.