Release Date: 8/24/2026
Format: WAV
Size: 4.35 GB
This dataset contains Sheng audio and accurate human-transcript pairs within the mobile money domain. It covers a wide range of demographics with a rich distribution across gender/age group pairs. Audio was collected from individual respondents in-field (with phone microphone as the most common recording methodology). Speakers were asked to answer prompts/questions that focused specifically on mobile money usage, and provided free-form, unstructured responses. Rather than reading predetermined sentences, respondents spoke naturally about their experiences and usage. This methodology captures a diverse range of speakers, utterances, speaking styles, and recording environments. This makes the dataset ideal for the development of ASR modelling and applications where code-switching between Sheng, Swahili and English is expected, and speech is likely to occur in real-world environments where there is high acoustic variability. A 1 hour sample is available for free download; the larger corpus of data (50 hours) is published under the paywall.
Pricing details
Your purchase is the license to the raw data. Once purchased, you're responsible for storing and using this dataset.
Your purchase includes a license to the data, paid directly to the dataset vendor, and a 5% ($500.00) platform fee paid to MDC.
Read More About The PoliciesThis dataset cannot be used to train any models or be used for applications that fit the following criteria:
Users must follow the data license as a governing usage guideline.
You agree not to attempt to determine the identity of speakers in this dataset Any attempt to clone the voice or train models that imitate the speakers in this dataset is forbidden.
Intended Use
This dataset can be used for a broad set of applications, but is mainly intended for building robust ASR systems that support Sheng.
Sheng is an urban language variety widely spoken in Kenya, particularly among younger populations. It blends Swahili and English with vocabulary and influences from other Kenyan languages.
Sheng is characterized by:
The speakers selected to contribute to this dataset are from rural Kenya and self-qualified as Sheng speakers.
The demographic distribution in the full dataset is:
| Gender | Unique Speakers | Recordings | Audio Hours |
|---|---|---|---|
| Men | 466 | 3,958 | 31.90 |
| Women | 285 | 2,321 | 17.99 |
| Undisclosed gender | 15 | 97 | 0.81 |
| Total | 766 | 6,376 | 50.70 |
| Age Group | Number of Recordings | Audio Hours |
|---|---|---|
| Ages 18–24 | 4,471 | 35.60 |
| Ages 25–34 | 1,558 | 12.60 |
| Ages 35–44 | 44 | 0.12 |
| Ages 45–54 | 67 | 0.75 |
| Ages 55–64 | 2 | 0.01 |
| Ages 65+ | 10 | 0.02 |
| Undisclosed age group | 224 | 1.60 |
| Total | 6,376 | 50.70 |
| Demographic Category | Number of Unique Speakers |
|---|---|
| Men | 466 |
| Women | 285 |
| Undisclosed gender | 15 |
| Total Unique Speakers | 766 |
| Age Group | Unique Speakers |
|---|---|
| Ages 18–24 | 541 |
| Ages 25–34 | 180 |
| Ages 35–44 | 6 |
| Ages 45–54 | 8 |
| Ages 55–64 | 1 |
| Ages 65+ | 1 |
| Unknown age group | 29 |
| Total | 766 |
Note: Age and gender categories are independent demographic attributes, so their counts should not be summed together. Certain respondents did not consent to demographic collection, hence the unknown values.
The dataset contains the following files and directories:
dataset/
├── audio/
│ └── *.wav
├── metadata.csv
└── README.md
audio/ — Contains the full WAV audio recordings.README.md — Provides basic documentation for the dataset.metadata.csv — Contains the transcriptions and associated metadata for each recording.In addition to the audio recordings and transcripts, the dataset includes metadata fields describing speaker characteristics and recording quality.
| Metadata Field | Description |
|---|---|
file_name | Relative path to the corresponding WAV audio file in the audio/ directory. |
transcription | Human-generated transcription of the audio recording, preserving the original wording, punctuation, and code-switching between Sheng, Swahili, and English. |
normalized_transcript | Normalized version of the transcription for automatic speech recognition (ASR), with standardized capitalization, punctuation, and whitespace while preserving the spoken words and language variations. |
speaker_age_range | Age range of the speaker, grouped into standardized age brackets. May contain Unknown when unavailable. |
speaker_gender | Gender of the speaker (Male, Female, or Unknown when unavailable). |
Audio was collected using mobile phone microphones through a structured data pipeline that requested WhatsApp voice notes from respondents. This approach was designed to facilitate participation among hard-to-reach populations.
Participants self-selected to take part in the collection process.
Participants were provided with a series of prompts and asked to record free-form spoken responses. Example prompts included:
Tell me about the last time you had a problem with your mobile service. What happened, and what did you do to try to fix it?
Tell me about the last time you contacted your mobile provider's customer service, such as Safaricom or Airtel. What was the problem, how did you contact them, and what happened?
After collection, submitted recordings were stored and reviewed for quality and submission accuracy.
Community experts in Sheng reviewed and transcribed each voice note. As a result, the dataset includes accurate transcription coverage that incorporates the full suite of code-switching and general language complexity.
Transcriptions were also normalized in normalized_transcript under the following rules:
Note: Numbers were not normalized and are mainly present in the decimal form, but may show up as words. Downstream implementation can decide which should be source of truth.
To improve transcription accuracy and completeness, a cleaning pipeline was developed to perform several quality checks:
Semantic check Determines whether the user's transcription is coherent and meaningfully answers the original prompt.
Syntax check Reviews spelling, capitalization, grammar, and punctuation.
Audio alignment Uses a combination of human review and AI-generated transcription to check the submitted transcription for general alignment with the source audio and overall accuracy.
Following these checks, revisions were made where necessary to improve the completeness and quality of the final transcriptions.
You might also like...