Datasets:

Kennethdot
/

Ghana_English-Twi_Code-switching_ASR

license: mit
dataset_info:
  features:
    - name: speaker_id
      dtype: string
    - name: age_range
      dtype: string
    - name: gender
      dtype: string
    - name: prompt_set
      dtype: string
    - name: transcript
      dtype: string
    - name: duration
      dtype: float32
    - name: split
      dtype: string
    - name: audio
      dtype: audio
    - name: file_name
      dtype: string
    - name: error
      dtype: string
  splits:
    - name: train
      num_bytes: 3440028114.5242105
      num_examples: 46303
    - name: validation
      num_bytes: 145645851.14104894
      num_examples: 1893
    - name: test
      num_bytes: 124887196.19536817
      num_examples: 1681
  download_size: 3793641295
  dataset_size: 3710561161.8606277
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
task_categories:
  - automatic-speech-recognition
language:
  - en
  - tw

Dataset Card for KasaSpeech

Dataset Description

Homepage: https://huggingface.co/datasets/Kennethdot/Ghana_English-Twi_Code_switching_ASR
License: CC BY 4.0
Languages: English (en), Twi/Akan (tw), English–Twi code-switching
Tasks: Automatic Speech Recognition (ASR), multilingual speech modeling, code-switching research

KasaSpeech is a large-scale English–Twi code-switching speech dataset created to advance speech AI research for low-resource African languages. The dataset contains transcribed speech recordings featuring natural switching between English and Twi, collected from diverse speakers primarily in Ghana.

Dataset Structure

Data Splits

The current version of the dataset shows these statistics

Split	Samples
Train	48,292
Test	1,005
Validation	581
Total	49,878

Data Fields

Field	Type	Description
`speaker_id`	string	Anonymous speaker identifier
`gender`	string	Speaker gender
`age_group`	string	Speaker age group
`audio`	Audio	Speech recording (48kHz)
`transcript`	string	Human-annotated transcript
`split`	string	Dataset split: train / validation / test

Example

from datasets import load_dataset, Audio

ds = load_dataset("Kennethdot/Ghana_English-Twi_Code_switching_ASR")
ds = ds.cast_column('audio', Audio(sampling_rate=16000))

sample = ds["train"][0]
print(sample["transcript"])  # e.g. "Ghanaian Jollof and Nigerian Jollof are both nice, nti gyae saa comparison no."
sample["audio"]                 # {"array": [...], "sampling_rate": 16000}

Data Collection

Speech recordings were voluntarily contributed by speakers and manually transcribed following unified annotation guidelines for English–Twi code-switching speech. Audio was cleaned, validated, and cast to 16kHz mono for speech modeling tasks.

Intended Use

KasaSpeech is intended for:

Automatic Speech Recognition (ASR) — training and evaluating ASR models on code-switched speech
Multilingual and code-switching research — studying intra-sentential and inter-sentential switching
African language technology development — building tools for Twi and Ghanaian English speakers
Speech representation learning — pre-training and fine-tuning models such as Whisper and wav2vec 2.0

Limitations

Demographic imbalance may exist across gender and age groups
Recording quality varies across speakers and environments
Primarily reflects Ghanaian English–Twi speech patterns and may not generalize to other Twi dialects

Citation

If you use KasaSpeech in your research, please cite:

@dataset{kasaspeech2026,
  title   = {KasaSpeech: An English--Twi Code-Switching Speech Dataset},
  author  = {Dotse, Kenneth},
  year    = {2026},
  url     = {https://huggingface.co/datasets/Kennethdot/Ghana_English-Twi_Code_switching_ASR}
}

Contact

For questions or collaboration, please open a discussion on the dataset page