KenSpeech: A Swahili Speech Dataset for ASR

Dataset Description

KenSpeech is a comprehensive Swahili speech dataset containing both read and spontaneous speech recordings from native Swahili speakers in Kenya. This dataset is designed for training and evaluating automatic speech recognition (ASR) and speech-to-text (STT) systems for Swahili.

Dataset Statistics

Metric	Value
Total Duration	27 hours 31 minutes 50 seconds
Read Speech Duration	26 hours 32 minutes 37 seconds
Spontaneous Speech Duration	59 minutes 13 seconds
Total Speakers	26
Female Speakers	19
Male Speakers	7
Lexicon Words	31,728+

Audio Format

Property	Value
Sampling Rate	16 kHz
Channels	Mono

---

Dataset Format

The dataset is distributed as Parquet files with embedded audio for optimal compatibility:

• Format: Apache Parquet (with embedded audio bytes)
• Encoding: UTF-8 for text fields
• Compatibility: Works with datasets 4.0.0+ without custom loading scripts

Data Fields

Column	Type	Description
audio	Audio	Audio waveform (decoded array + sampling_rate)
source_folder	string	Origin folder (stt_dictionary or stt_transcripts)
gender	string	Speaker gender (male or female)
speaker	string	Speaker identifier (speaker_1, speaker_2, etc.)
transcript	string	Transcription text

Example Record

{
    'audio': {'path': '...', 'array': array([0.001, -0.003, ...]), 'sampling_rate': 16000},
    'source_folder': 'stt_dictionary',
    'gender': 'female',
    'speaker': 'speaker_1',
    'transcript': 'masaa mawili kabla basi kuwasili...'
}

---

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Kencorpus/KenSpeech")

# Access a sample
sample = dataset['train'][0]
print(sample['transcript'])
print(sample['gender'])
print(sample['speaker'])
print(sample['audio']['sampling_rate'])  # 16000
print(sample['audio']['array'].shape)    # audio waveform

Filtering by Gender

from datasets import load_dataset

dataset = load_dataset("Kencorpus/KenSpeech")

# Get female speakers only
female_data = dataset['train'].filter(lambda x: x['gender'] == 'female')
print(f"Female samples: {len(female_data)}")

# Get male speakers only
male_data = dataset['train'].filter(lambda x: x['gender'] == 'male')
print(f"Male samples: {len(male_data)}")

Training an ASR Model

from datasets import load_dataset
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

# Load dataset
dataset = load_dataset("Kencorpus/KenSpeech")

# Load a multilingual model
model_name = "facebook/wav2vec2-large-xlsr-53"
processor = Wav2Vec2Processor.from_pretrained(model_name)
model = Wav2Vec2ForCTC.from_pretrained(model_name)

# Process a sample
sample = dataset['train'][0]
inputs = processor(sample['audio']['array'], sampling_rate=16000, return_tensors="pt")

---

Additional Resources

Pronunciation Lexicon (lexicon.csv)

A Swahili lexicon-phone dictionary with over 31,000 words and their phonetic transcriptions.

Format: word,phoneme_sequence

wanapaswa,W AH N AH P AH S W AH
wanasema,W AH N AH S EH M AH
wanataka,W AH N AH T AH K AH

Transcript-only Data (transcripts_only.csv)

Additional transcripts from the stt_transcripts collection without corresponding audio.

---

Speech Types

Type	Duration	Percentage
Read Speech	26h 32m 37s	96.4%
Spontaneous Speech	59m 13s	3.6%

---

Intended Uses

• Training automatic speech recognition (ASR) systems for Swahili
• Evaluating speech-to-text models
• Phonetic and linguistic research on Swahili
• Building text-to-speech (TTS) systems
• Transfer learning for other Bantu languages

---

Dataset Curators

• Dorcas Awino
• Dr. Benard Okal
• Khalid Kitito
• Owiny Japheth Otieno

---

Citation

@article{wanjawa2022kencorpus,
  title={Kencorpus: A Kenyan Language Corpus of Swahili, Dholuo and Luhya for Natural Language Processing Tasks},
  author={Wanjawa, Barack W. and Wanzare, Lilian D. and Indede, Florence and McOnyango, Owen and Ombui, Edward and Muchemi, Lawrence},
  journal={arXiv preprint arXiv:2208.12081},
  year={2022}
}

---

Links

• Research Paper: https://arxiv.org/abs/2208.12081
• Dataverse: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/KLCKL5

---

License

This dataset is licensed under CC-BY-4.0.

---

Acknowledgments

This dataset is part of the Kencorpus project, which aims to create NLP and speech resources for low-resource Kenyan languages.