This repository contains an Afrikaans simulated telephone speech dataset with segmented audio and fine-grained transcripts. Each row in the metadata corresponds to a short segment from a longer call, with speaker information and timing metadata. The data can be used for automatic speech recognition (ASR), speaker-aware ASR, and related research on Afrikaans conversational speech.

Dataset Structure

Segmented audio: Channel-separated utterances (short clips) stored as .wav files under data/audio/segmented/<domain>/ (e.g. data/audio/segmented/debtcollection/af_za_debtcollection_002-part-1.wav). These are the clips referenced by transcripts.csv and exposed via the audio feature in the dataset script.
Full-length audio: Complete, channel-separated call recordings stored as .wav files under data/audio/full_length/<domain>/ (e.g. data/audio/full_length/debtcollection/af_za_debtcollection_002.wav). These files are included in the repository for users who want to work at the full-call level.
Transcripts and segment metadata: data/transcripts/transcripts.csv, containing one row per audio segment with the segment transcript and associated metadata (call ID, speaker ID, channel, start/end times, duration).
Call and speaker metadata: metadata/call_metadata.csv, containing call- and speaker-level information such as gender, age range, home language, education level, and topic.

The dataset is organised so that the paths referenced in transcripts.csv use relative paths from the dataset root to reference the corresponding segmented audio files.

Dataset Statistics

This repository contains the Afrikaans 50h ASR Dataset developed by Way With Words.

Source corpus (Way With Words product):

Hours: 50 hours of conversational speech.
Speakers: 46 recorders.
Download size: 38 GB (WAV).
Age range: 18–69; distribution: 18–29 (9 recorders), 30–49 (28 recorders), 50–69 (9 recorders).
Gender Split: 26 women, 20 men.
Hours by domain: Retail 12:21:57, Debt Collection 12:19:33, Insurance 12:19:23, Travel 13:02:28.

This repository (from metadata/call_metadata.csv):

Full-length calls: 222 calls (each with 2 channels).
Channel-speaker rows: 444 (one row per call × channel × speaker in the metadata).
Age ranges: 18–29, 30–49, 50–69.
Home language: Afrikaans (all speakers in this release).
Education levels: High School, Diploma, Undergraduate, Graduate.
Geographic origin: multiple South African provinces (e.g. Gauteng, KwaZulu-Natal, Western Cape, Mpumalanga, Northern Cape, Eastern Cape).
Recording devices: laptops and desktop computers.

Data Fields (Transcripts)

The main transcription manifest data/transcripts/transcripts.csv contains the following columns:

file_name
segment
channel
duration
speaker
start
end
transcript

For use with the 🤗 datasets library, these map conceptually to the following logical fields:

audio: path to the segment audio file (taken from segment, possibly with a directory prefix when loading).
sentence: transcript text for ASR (taken from transcript).
speaker_id: speaker identifier string (taken from speaker, e.g. spkr_07).
call_id: identifier for the full-length call (taken from file_name, e.g. af_za_debtcollection_002.wav).
channel: channel ID within the call (taken from channel, e.g. chan_1 / chan_2).
start: segment start time in seconds within the full call (taken from start).
end: segment end time in seconds within the full call (taken from end).
duration: segment duration in seconds (taken from duration).

The dataset loading script keeps the original CSV column names but exposes these logical fields through the datasets.Features definition so that downstream users can access them in a consistent way.

Call and Speaker Metadata

The file metadata/call_metadata.csv contains call- and speaker-level metadata for each channel in the full-length calls. Key columns include:

File Name – identifier for the full-length call (e.g. af_za_debtcollection_002, without the .wav extension).
Channel – channel ID within the call (1 or 2).
Speaker ID – ID of the recorder/speaker on that channel (e.g. spkr_07).
Gender – gender category of the speaker (e.g. M, F).
Age – age range bucket (e.g. 18 - 29, 30 - 49).
Home Language – primary language of the speaker (e.g. Afrikaans).
Education Level – highest completed education level (e.g. High School, Undergraduate, Graduate).
Place of Origin – region or province associated with the speaker (e.g. Gauteng, KwaZulu-Natal).
Topic / Scenario – high-level description of the call content (e.g. Debt Collection, Payment request).
Recording Device – device type used for recording (e.g. Laptop, Desktop Computer).

This file is not directly loaded by the dataset script but can be joined to the main dataset using the call identifier (file_name / File Name with or without the .wav extension), channel, and speaker ID for analyses that require demographic or call-level context.

Example Call Metadata Row

An example row from metadata/call_metadata.csv looks like:

File Name: af_za_debtcollection_002
Channel: 1
Speaker ID: spkr_07
Gender: M
Age: 50 - 69
Home Language: Afrikaans
Education Level: Graduate
Place of Origin: Gauteng
Topic / Scenario: Debt Collection / Payment request
Recording Device: Laptop

Transcription and Annotation Guidelines

This repository includes TRANSCRIPTION_GUIDE.md, which describes in detail the conventions used for transcription and annotation (e.g. handling of hesitations, fillers, numbers, and non-speech events). Users who need a deeper understanding of the transcript format or who plan to post-process the text are encouraged to consult this guide.

Supported Tasks

Automatic Speech Recognition (ASR)
- Input: audio
- Target: sentence
Speaker-aware ASR / Speaker Diarization Research
- Input: audio
- Targets: sentence, speaker_id, channel, call_id, start, end

Data Splits

All segments are currently stored in a single transcription manifest file (e.g. data/transcripts/transcripts.csv). Consumers of the dataset can create their own train/validation/test splits by filtering on call IDs, speakers, or other criteria.

License

This dataset is made available by Way With Words Limited / Way With Words SA (Pty) Ltd under the terms of the Way With Words Speech Collection Dataset Licence Agreement.

The full, legally binding text of the Licence Agreement is published at:
https://waywithwords.net/legal/speech-collection-dataset-licence-agreement/

This repository includes a LICENSE file which summarises the key points of the Licence Agreement for convenience only. In the event of any inconsistency, the Licence Agreement at the URL above prevails.

Usage with 🤗 Datasets

This dataset can be loaded from the Hugging Face Hub as follows:

from datasets import load_dataset

ds = load_dataset("waywithwords/www-za-afr-cx", split="train")

example = ds[0]
audio = example["audio"]
text = example["sentence"]
speaker = example["speaker_id"]

This version of the dataset does not define predefined splits; use split="train" (or omit the split argument to load the default split) and create your own splits as needed.

Example Segment

A typical example from the dataset looks like:

audio: segmented/af_za_debtcollection_002-part-1.wav
sentence: "Goeie middag."
speaker_id: spkr_07
call_id: af_za_debtcollection_002.wav
channel: chan_1
start / end (s): 5.280 → 7.855
duration (s): 2.575

Ethical Considerations

This dataset consists of simulated telephone conversations. Any names, account details, personal information, company names, brand names, product names, or other identifying details appearing in the audio or transcripts are used solely in a simulated context and do not imply any affiliation with, endorsement by, or representation of any real individual, organisation, product, or service.

While this substantially reduces privacy risks compared to real-world recordings, users should still:

Ensure their use of the dataset complies with the Licence Agreement and any applicable laws or internal policies.
Users should be aware that this dataset contains simulated rather than real conversational data when interpreting model behaviour.

Downloads last month: 10

Total file size:

6.88 MB