AfriSwitchCare: African Code-Switched Clinical Conversation Benchmark

Dataset Description

AfriSwitchCare is a test benchmark of simulated doctor–patient medical conversations in which an African language is code-switched with English. It spans 8 languages — Amharic, French, Hausa, Kinyarwanda, Pidgin, Swahili, Igbo, Yoruba. Six languages (Amharic, French, Hausa, Kinyarwanda, Pidgin, Swahili) cover the same 12 clinical conditions (12 conversations each); Igbo (5) and Yoruba (6) are partial. 83 conversations in total (8.95 hours).

This release is distributed as a single test split (evaluation-only benchmark).

• License: CC BY NC SA 4.0
• Conversations: 83
• Total duration: 8.95 hours
• Split: test (only)

Clinical conditions (12)

Acute Appendicitis, Asthma, Bronchopneumonia, Diabetes Mellitus, Febrile Convulsion, Hypertension, Osteoarthritis, Pelvic Inflammatory Disease (PID), Stroke, Tuberculosis, Depression, Drug-induced Psychosis. (Igbo and Yoruba cover a subset of these.)

Dataset Statistics

Language	Conversations	Diagnoses	Avg. Turns	Avg. Duration	Hours	Avg. CMI	Avg. Switch Pts
Amharic	12	12	42.8	53.0 min	0.88	35.6	66.6
French	12	12	49.5	78.5 min	1.31	30.5	39.8
Hausa	12	12	46.2	41.3 min	0.69	35.5	57.8
Kinyarwanda	12	12	71.2	93.5 min	1.56	26.6	55.0
Pidgin	12	12	79.9	86.0 min	1.43	8.4	29.6
Swahili	12	12	N/A	92.3 min	1.54	37.4	96.4
Igbo	5	5	127.0	39.7 min	0.66	27.8	156.8
Yoruba	6	5	96.0	52.5 min	0.88	34.1	76.8
Total	83				8.95

Turns = number of speaker turns ([Speaker N]). Swahili transcripts carry no speaker markers, so turns are N/A for Swahili. CMI (Code-Mixing Index) and switch points follow Das and Gambäck (2014) / Gambäck and Das (2016).

Dataset Structure

Each example is one conversation:

• audio: the conversation audio, 16 kHz mono HuggingFace Audio feature (bytes + path)
• language: matrix language
• diagnosis: the clinical condition
• transcription: verbatim human transcription (plain text)
• transcription_tagged: same transcription with English spans wrapped in [[EN]]…[[/EN]]
• num_turns: number of speaker turns (null where speaker markers are absent)
• cmi: Code-Mixing Index
• num_switch_points: language-alternation count
• duration: seconds

Each language is a separate config with a single test split.

Dataset Creation

Source Data

Simulated doctor-patient consultations recorded for clinical-domain code-switch ASR. Bilingual annotators produced verbatim transcriptions of the conversations.

The released dataset contains transcriptions and processed audio only; no patient data is involved (consultations are simulated) and no annotator demographic information is included.

Licensing Information

Released under Attribution-NonCommercial-ShareAlike 4.0 (CC BY NC SA 4.0).