onos9/tcnspeech_sermon_corpus_24h

TCNSpeech: A Community-Curated Speech Corpus for Sermons

A multispeaker sermon corpus for Automatic Speech Recognition (ASR) in Nigerian English

Overview

TCNSpeech is an open community-curated multispeaker speech corpus of English sermons from Nigerian preachers. This dataset captures domain-specific sermon content with Nigerian English accents, curated by volunteers from the Digital Tech Community Group at The Covenant Nation (TCN) in Lagos, Nigeria. The dataset demonstrates the power of community-driven data annotation for building speech recognition resources in underrepresented accents and domains.

Paper: TCNSpeech: A Community-Curated Speech Corpus for Sermons (AfricaNLP workshop at ICLR 2022)

Dataset Details

Statistics

• Total Audio Duration: 24 hours
• Total Transcribed Samples: 8,550 audio clips
• Male Speakers: 4,950 samples (~13 hours)
• Female Speakers: 3,600 samples (~11 hours)
• Language: English (Nigerian English accent)
• Domain: Religious sermons
• Geographic Context: Primarily speakers from Lagos, southwestern Nigeria

Audio Specifications

• Format: WAV
• Sample Rate: 16,000 Hz
• Chunk Duration: 10 seconds per file
• Preprocessing: Audio files chunked using Audacity

Dataset Structure

tcnspeech_sermon_corpus_24h/
├── Male/
│   ├── audios/
│   │   └── [2,128 audio files] (~13 hours, 4,950 samples)
│   └── transcripts.txt
├── Female/
│   ├── audios/
│   │   └── [1,044 audio files] (~11 hours, 3,600 samples)
│   └── transcripts.txt
├── TCNSpeech Dataset_24h - male_.csv
├── TCNSpeech Dataset_24h - female_.csv
└── README.md

Each transcription line in the transcript files corresponds to an audio file of the same name (without the .wav extension). CSV Files:

TCNSpeech Dataset_24h - female_.csv (694 KB) - Maps 3,600 female speaker transcripts to audio files TCNSpeech Dataset_24h - male_.csv - Maps 4,950 male speaker transcripts to audio files

File Mapping with CSV Files The dataset includes CSV files that explicitly map each transcript to its corresponding audio file for both gender splits. Each CSV file contains two columns: transcript (the full text transcription of the audio) and audio_file_path (the path to the corresponding WAV audio file). Example CSV entry from female split: "Shola Salako said something she said it's not the one you do here...",TCNSpeech Dataset_24h/female/audios/BE_2071 CSV Files: TCNSpeech Dataset_24h - female_.csv (694 KB) maps 3,600 female speaker transcripts to their corresponding audio files. TCNSpeech Dataset_24h - male_.csv maps 4,950 male speaker transcripts to their corresponding audio files. This structured mapping ensures precise alignment between audio files and their transcriptions, making it easy to load corresponding pairs for speech recognition tasks.

Dataset Content

The dataset includes typical church-related experiences and sermon elements:

• Sermon speech: Standard preaching and teaching content
• Special vocal elements: Praying, singing, speaking in tongues
• Acoustic phenomena: Clapping, music interludes, congregation responses
• Linguistic content: Biblical terms, names, locations, bible chapter references, church-specific vocabulary

Transcription Conventions

The transcriptions follow detailed annotation guidelines: 3https://bit.ly/audio_data_transcription_guide

• Uniform spelling: Full spellings (e.g., "verses" not "vs", "Corinthians" not "Cor.")
• Numbers as figures: All numbers, including bible chapters/verses written as numerals (e.g., "James 2 1" not "James 2 verses 1")
• Special markers:
  • [music] - for songs or instrumentals
  • [unknown] - for unclear spoken words
  • [clap] - for clapping sounds
  • [speaking in tongues] - for glossolalia
  • [prayer] - for prayer segments

Recommended Use Cases

This dataset is particularly useful for:

• Nigerian English ASR: Training accent-specific speech recognition models
• Domain-specific Speech Recognition: Sermon/religious content transcription
• Multilingual/Multi-accent Research: Understanding Nigerian English phonetics and prosody
• Low-resource Language Technology: Community-driven data annotation methodologies
• Church/Religious Institution Applications: Real-time sermon transcription systems

Dataset Characteristics

Accent and Dialect

The dataset captures Nigerian English as spoken by preachers in Lagos and surrounding areas. This includes characteristic Nigerian English intonation, inflection patterns, and pronunciation influenced by local languages and speaking traditions.

Challenges and Features

• Speaking style variations: Sermon-specific speaking patterns (emphatic delivery, emotional expression)
• Code-switching potential: English with religious terminology and biblical references
• Acoustic diversity: Range of speaker voice characteristics, ages, and speech rates
• Real-world conditions: Actual church recordings including congregation sounds (managed through careful speaker selection)

Paper Results

The paper includes experimentation with two speech recognition models trained on TCNSpeech using NVIDIA NeMo's QuartzNet 15x5 ASR architecture:

Experiment	Sermon Data	en-ng Data	Total Duration	Train	Validation	Validation WER
Model 1	4.43 hours	5.77 hours	10.20 hours	9.95	0.25	0.35%
Model 2	13.48 hours	5.77 hours	19.25 hours	19.00	0.25	0.31%

Best performing model achieved a Word Error Rate of 0.31% when trained on the larger dataset combined with the Nigerian English multi-speaker speech dataset.

Data Curation Methodology

TCNSpeech demonstrates a community-focused data curation approach:

1. Community Engagement: 71 volunteers from the Digital Tech Community Group at The Covenant Nation
2. Preparation: Detailed concept note and annotation guide for volunteer transcribers
3. Tools: Google Drive (storage), Google Forms (registration), Google Sheets (tracking), Audacity (preprocessing)
4. Quality Assurance: Double-checked transcriptions for alignment with guidelines and filename correctness
5. Support: Real-time support through WhatsApp groups and email

How to Load the Dataset

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("Wuraola/tcnspeech_sermon_corpus_24h")

# Load specific splits
male_data = load_dataset("Wuraola/tcnspeech_sermon_corpus_24h", split="male")
female_data = load_dataset("Wuraola/tcnspeech_sermon_corpus_24h", split="female")

For a detailed tutorial on loading and working with the dataset, see the sample notebook.

Licensing and Attribution

License: MIT License

Citation

If you use TCNSpeech in your research, please cite:

@inproceedings{oyewusi2022tcnspeech,
  title={TCNSpeech: A Community-Curated Speech Corpus for Sermons},
  author={Oyewusi, Wuraola Fisayo and Ibejih, Sharon and Uzomah, Soromfe and Joseph, Elizabeth and others},
  booktitle={AfricaNLP workshop at ICLR 2022},
  year={2022}
}

Contributors

The dataset was created through the collaborative efforts of the Digital Tech Community Group at The Covenant Nation (TCN) in Lagos, Nigeria, including:

Wuraola Fisayo Oyewusi, Sharon Ibejih, Soromfe Uzomah, Elizabeth Joseph, Cynthia Olawuyi, Folakunmi Ojemuyiwa, Benedicta Johnson-Onuigwe, Omolola Taiwo, Akintunde Akinpelumi, Olabisi Adesina, Ayodele Noutouglo, Adeola Adeoba, Andrew Akoh, Chukwuemeka Nwachukwu, Opeyemi Agbabiaje, Itunu Falade, Olukemi Erhunmwunsee, Oluwatobiloba Dada, Oluwatobi Osibeluwo, Ehis Akene, Udim Akpan, Moira Amadi-Emina, Jaiyeola Marquis, Michael Senapon Bojerenu, Gbolahan Olumade, Oluwagbemi Lesi, Timothy Ezeh, Oluwadamilola Oguntoyinbo, Tosan Mogbeyiteren, Felicia Oresanya, Samuel Chika, and Sodiq Akinjobi.

Dataset Availability

The dataset is made publicly available as a contribution to speech recognition research and the broader scientific community. It represents an important resource for developing speech technology for underrepresented languages and dialects.

Limitations and Known Issues

The dataset may contain transcription errors despite quality assurance efforts. Speakers are primarily from Lagos and southwestern Nigeria, which may limit generalization to other Nigerian regions. Some acoustic challenges from real-world church recordings remain. The dataset is optimized for sermon content and may not generalize equally to other speech domains.

Contact

For questions or inquiries about the dataset, please contact the Digital Tech Community Group at The Covenant Nation: ccgdigitaltech@gmail.com

---

Last Updated: November 2025
Dataset Version: 1.0