AfriSpeeches: Official Speeches of African Heads of State

AfriSpeeches is a dedicated speech research corpus preserving, analyzing, and attributing the oratorical voices and official addresses of African heads of state, presidents, prime ministers, and historical leaders across the African continent.

Spanning 51 African nations in English, French, Portuguese, and Arabic, this dataset is created to empower speech research—enabling academic study of African acoustic phonetics, regional oratorical accents, multilingual speech-to-text (ASR), and computational political discourse, while giving full attribution to the statesmen whose voices helped shape modern continental history.

---

Attribution & Provenance

Every audio segment in AfriSpeeches is attributed to the specific statesman who delivered it:

• Full Attribution: Every record includes the leader's full name, country of state, national ISO code, spoken language, speech title, and the verifiable source recording URL.
• Natural Oratorical Chunking: Rather than slicing audio at arbitrary time boundaries, speeches are split strictly at natural pauses identified via Silero Voice Activity Detection (median segment duration of 6.5 minutes, capped at <= 8.0 minutes). This preserves complete oratorical thoughts, rhetorical delivery, and natural cadence.
• Verbatim Transcriptions: Paired with complete speech-to-text transcriptions verified for acoustic continuity and textual integrity.
• Embedded Audio: Audio is encoded in 24 kHz mono FLAC embedded directly inside browser-streamable Parquet shards, eliminating the need for complex extraction or separate multi-gigabyte archive downloads.

---

Dataset Summary

• Total Speech Audio: 269.0 hours of verified speech
• Total Speech Segments: 2,757 natural-pause segments
• Sovereign Nations Represented: 51 African countries
• Attributed Heads of State & Leaders: 139 statesmen
• Languages: ar, en, fr, pt (English, French, Portuguese, Arabic)
• Audio Specification: 24,000 Hz, 16-bit mono FLAC embedded inline

---

Quickstart & Usage

You can load and stream AfriSpeeches directly via the Hugging Face datasets library:

from datasets import load_dataset, Audio

# Stream directly from Hugging Face Hub (no massive disk download required)
dataset = load_dataset("AfriSpeech/afrispeeches", split="train", streaming=True)
dataset = dataset.cast_column("audio", Audio(decode=False))

for sample in dataset:
    speaker = sample["speaker"]
    country = sample["country"]
    lang = sample["language"]
    duration = sample["duration"]
    transcript = sample["text"]
    url = sample["url"]
    
    print(f"[{country} ({lang})] {speaker} ({duration:.1f}s):")
    print(f"  Transcript: {transcript[:120]}...")
    print(f"  Source:     {url}\n")
    break

Loading Audio Waveforms into PyTorch / Torchaudio

import io
import soundfile as sf
from datasets import load_dataset, Audio

dataset = load_dataset("AfriSpeech/afrispeeches", split="train", streaming=True)
dataset = dataset.cast_column("audio", Audio(decode=False))
sample = next(iter(dataset))

# Read FLAC bytes directly into a float32 numpy array / PyTorch tensor
waveform, sample_rate = sf.read(io.BytesIO(sample["audio"]["bytes"]), dtype="float32")
print("Speaker:    ", sample["speaker"])
print("Sample Rate:", sample_rate, "Hz")
print("Samples:    ", waveform.shape)

---

Dataset Schema

Each row contains clean, standardized fields designed for speech research and model training:

Field	Type	Description
audio	dict	Embedded audio with bytes (24 kHz FLAC) and filename path
text	string	Verbatim speech-to-text transcript
speaker	string	Full name of the African head of state or political leader
country	string	Country of the speaker (e.g. Ghana, Nigeria, Senegal, Egypt)
iso3	string	ISO 3166-1 alpha-3 nation code (e.g. GHA, NGA, SEN, EGY)
language	string	Spoken language code (en, fr, pt, ar)
title	string	Title, address, or occasion of the official speech
duration	float32	Segment duration in seconds (median 6.5 min, <= 8.0 min)
url	string	Source recording URL

---

Coverage by Country & Featured Leaders

Country	Segments	Audio Hours	Featured Leaders
Benin	151	15.1h	Patrice Talon, Romuald Wadagni, ...
Guinea	148	14.6h	Alpha Condé, Mamady Doumbouya, ...
Democratic Republic of the Congo	128	12.6h	Félix Antoine Tshisekedi Tshilombo, Joseph Kabila Kabange, ...
Niger	123	12.3h	Abdourahamane Tchiani, Mahamadou Issoufou, ...
South Africa	123	12.2h	Jacob Gedleyihlekisa Zuma, Kgalema Motlanthe, ...
Chad	116	11.5h	Hissène Habré, Idriss Déby Itno, ...
Angola	112	11.2h	António Agostinho Neto, José Eduardo dos Santos, ...
Senegal	110	10.7h	Abdoulaye Wade, Bassirou Diomaye Faye, ...
Republic of the Congo	92	9.3h	André Milongo, Denis Sassou Nguesso, ...
Cabo Verde	84	8.0h	Jorge Carlos Fonseca, José Maria Neves, ...
Liberia	81	7.9h	Ellen Johnson Sirleaf, George Weah, ...
Kenya	80	7.9h	Mwai Kibaki, Uhuru Muigai Kenyatta, ...
Morocco	68	6.5h	Aziz Akhannouch, King Mohammed VI, ...
Namibia	66	6.4h	Hage Gottfried Geingob, Nangolo Mbumba, ...
Togo	63	5.7h	Faure Essozimna Gnassingbé, Gnassingbé Eyadéma, ...
Egypt	63	6.2h	Abdel Fattah el-Sisi, Hosni Mubarak, ...
Mali	61	5.8h	Amadou Toumani Touré, Assimi Goïta, ...
Ghana	60	5.8h	John Atta Mills, John Dramani Mahama, ...
Zimbabwe	60	6.0h	Canaan Sodindo Banana, Emmerson Dambudzo Mnangagwa, ...
Algeria	58	5.9h	Abdelaziz Bouteflika, Abdelmadjid Tebboune, ...
Mozambique	58	5.5h	Armando Guebuza, Daniel Francisco Chapo, ...
Somalia	57	5.2h	Hassan Sheikh Mohamud, Mohamed Abdullahi Mohamed, ...
Sudan	56	5.2h	Abdalla Hamdok, Abdel Fattah al-Burhan, ...
Gabon	52	5.2h	Ali Bongo Ondimba, Brice Clotaire Oligui Nguema, ...
Nigeria	52	5.1h	Bola Ahmed Tinubu, Goodluck Ebele Jonathan, ...
Malawi	52	5.4h	Arthur Peter Mutharika, Joyce Banda, ...
Tunisia	50	4.5h	Beji Caid Essebsi, Kaïs Saïed, ...
Zambia	43	4.4h	Edgar Chagwa Lungu, Hakainde Hichilema, ...
Sierra Leone	40	3.9h	Ernest Bai Koroma, Julius Maada Bio
Burkina Faso	39	4.2h	Ibrahim Traoré, Paul-Henri Sandaogo Damiba, ...
Mauritius	39	3.6h	Ameenah Gurib-Fakim, Dharambeer Gokhool, ...
Gambia	38	3.7h	Adama Barrow, Dawda Jawara, ...
Sao Tome and Principe	31	2.7h	Carlos Manuel Vila Nova, Evaristo do Espírito Santo Carvalho, ...
Tanzania	30	2.6h	Jakaya Mrisho Kikwete, John Pombe Joseph Magufuli, ...
Rwanda	29	3.0h	Juvénal Habyarimana, Paul Kagame
South Sudan	28	2.8h	Abel Alier, Riek Machar, ...
Burundi	28	2.7h	Domitien Ndayizeye, Pierre Nkurunziza, ...
Libya	25	2.3h	Abdoul Hamid Dbeibeh, Fayez al-Sarraj, ...
Lesotho	24	2.2h	Letsie III, Moeketsi Majoro, ...
Seychelles	24	2.4h	Danny Faure, Mathew Antonio Patrick Herminie, ...
Mauritania	16	1.6h	Mohamed Ould Abdel Aziz, Mohamed Ould Cheikh El Ghazouani
Cameroon	14	1.4h	Ahmadou Babatoura Ahidjo
Ethiopia	13	1.1h	Abiy Ahmed, Sahle-Work Zewde, ...
Comoros	13	1.3h	Ahmed Abdallah Mohamed Sambi, Azali Assoumani, ...
Guinea-Bissau	12	1.2h	Horta Inta-A Na Man, José Mário Vaz
Equatorial Guinea	11	1.1h	Francisco Macías Nguema, Teodoro Obiang Nguema Mbasogo
Madagascar	11	1.0h	Andry Rajoelina, Michael Randrianirina
Uganda	7	0.6h	Apollo Milton Obote, Tito Lutwa Okello, ...
Djibouti	7	0.7h	Ismaïl Omar Guelleh
Eritrea	6	0.6h	Isaias Afwerki
Eswatini	5	0.5h	Russell Mmiso Dlamini

---

Curation & Quality Assurance

1. Acoustic Standardization: Audio was normalized to 16-bit PCM mono with middle-50% speech extraction and peak stabilization.
2. Strict Quality Gates: Speech segments were evaluated via DNSMOS (sig_mean >= 3.0, sig_p10 >= 2.5) and minimum voice activity ratio (>= 0.15). Any audio corrupted by excessive stadium cheering, military parade music, or non-speech noise was rigorously excluded.
3. Natural Oratorical Splitting: Silero VAD identified speech silences to segment recordings without cutting mid-sentence, preserving complete rhetorical phrases.
4. ASR Continuity Auditing: Sub-windows were transcribed via Google ASR and audited to ensure complete textual continuity and eliminate decoder repetition or silent omissions.

---

Intended Applications

• African-Accented Speech Recognition (ASR): Benchmarking and fine-tuning acoustic models (Whisper, MMS, Conformer) on diverse regional African accents across official languages.
• Acoustic Phonetics & Dialectology: Quantitative research into the prosody, intonation, and phonetic features of prominent African political speakers.
• Computational Rhetoric & Political Science: Longitudinal study of African political discourse, pan-Africanist ideology, and diplomatic rhetoric.

---

License & Attribution

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0) license.

• Attribution: When using this dataset in academic research or applications, please cite the AfriSpeech / Ghana NLP community and cite the attributed speakers.
• Non-Commercial: Intended for non-commercial academic and linguistic research.