AfriSpeech/afrispeeches-aligned

AfriSpeeches Aligned: Enhanced & Aligned Speeches of African Heads of State

AfriSpeeches Aligned is a sentence-aligned, studio-enhanced speech corpus featuring African heads of state and leaders across 51 African nations, partitioned into dedicated train (85%) and validation (15%) splits.

Each continuous oratorical speech from AfriSpeeches has been:

1. Sentence-Aligned: Processed through CTC forced alignment (Meta MMS) into clean 3-to-15 second sentence and phrase utterances.
2. Audio-Enhanced: Processed through Resemble Enhance (deep neural denoising and diffusion bandwidth enhancement) to remove acoustic room reverb, microphone distortion, and environmental noise while preserving authentic speaker cadence and intonation.
3. Structured Hierarchy: Partitioned by Language -> Country -> Speaker with balanced train (85%) and validation (15%) splits across all leaders, and each leader published as its own Country_person config.

---

Dataset Summary

• Total Aligned Clips: 150,689 clips (128,464 train / 22,225 validation)
• Total Enhanced Audio: 214.9 hours (183.1h train / 31.7h validation)
• Countries Represented: 51 African nations
• Attributed Leaders: 139 heads of state
• Languages: ar, en, fr, pt
• Audio Specification: 24,000 Hz 16-bit mono FLAC

---

Loading Subsets (Configs)

Every leader is exposed as its own config named Country_person, containing that leader's own train and validation clips. The default config contains the full corpus.

For bilingual leaders (e.g. Arabic & English, or Arabic & French), dedicated language subsets are provided as Country_person_lang in addition to their overall profile:

from datasets import load_dataset, Audio

# 1. Full corpus (all leaders across Africa)
train_ds = load_dataset("AfriSpeech/afrispeeches-aligned", split="train", streaming=True)

# 2. A single monolingual leader
kenya = load_dataset(
    "AfriSpeech/afrispeeches-aligned",
    "Kenya_Uhuru_Muigai_Kenyatta",
    split="validation",
    streaming=True,
)

# 3. A bilingual leader in a specific language (e.g. Egypt's Hosni Mubarak in Arabic)
mubarak_ar = load_dataset(
    "AfriSpeech/afrispeeches-aligned",
    "Egypt_Hosni_Mubarak_ar",
    split="train",
    streaming=True,
)

# 4. Same leader in English
mubarak_en = load_dataset(
    "AfriSpeech/afrispeeches-aligned",
    "Egypt_Hosni_Mubarak_en",
    split="train",
    streaming=True,
)

---

Quickstart & Usage

Non-streaming loads work the same way; drop the config name for a full-corpus parquet download.

from datasets import load_dataset, Audio

full = load_dataset("AfriSpeech/afrispeeches-aligned", split="validation")
print(full)

---

## Schema

| Field | Type | Description |
|---|---|---|
| `audio` | `dict` | Embedded audio with `bytes` (24 kHz enhanced FLAC) and filename `path` |
| `text` | `string` | Aligned sentence/phrase transcript |
| `speaker` | `string` | Head of state or leader's name |
| `country` | `string` | Country of state |
| `iso3` | `string` | ISO 3166-1 alpha-3 country code |
| `language` | `string` | Language code (`en`, `fr`, `pt`, `ar`) |
| `split` | `string` | Partition split: `train` (85%) or `validation` (15%) |
| `duration` | `float32` | Utterance duration in seconds (3-15s) |
| `start_s` | `float32` | Offset start within original recording |
| `end_s` | `float32` | Offset end within original recording |
| `source_chunk_id`| `string` | Parent AfriSpeeches chunk identifier |
| `title` | `string` | Original speech title |
| `url` | `string` | Source recording URL |

---

## License & Attribution

Released under the **Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0)** license.

Developed by the **[AfriSpeech](https://huggingface.co/AfriSpeech)** / **[Ghana NLP](https://ghananlp.org)** community.