AfriSwitch: In-the-Wild African Code-Switched Speech Benchmark

Dataset Description

AfriSwitch is a 54.41-hour, human-transcribed test benchmark of in-the-wild, conversational code-switched speech spanning 14 African languages: Amharic, Pidgin, Kinyarwanda, Yoruba, Hausa, Oromo, Igbo, Zulu, French, Shona, Swahili, Tswana, Luganda, Afrikaans, each switching with English.

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

• License: CC BY NC SA 4.0
• Total duration: 54.41 hours
• Total utterances: 16,602
• Total code-switch events (S*): 69,928
• Split: test (only)

Dataset Statistics

Language	Hours	Utterances	Avg. Switch Points	Total S*	CMI
Kinyarwanda	5.00	1,577	4.51	7,108	16.95
Amharic	5.00	1,229	4.60	5,650	13.11
Zulu	5.00	1,465	4.40	6,447	24.76
Igbo	5.00	1,848	4.10	7,575	27.64
Yoruba	5.00	1,877	5.33	10,002	22.93
Hausa	5.00	1,515	4.00	6,053	13.09
Pidgin	4.56	1,801	1.46	2,621	4.19
Oromo	4.25	1,217	2.95	3,586	13.39
Swahili	3.89	650	10.29	6,689	25.72
Shona	3.86	1,155	4.85	5,599	24.55
French	3.22	903	3.00	2,713	10.92
Tswana	2.74	805	4.38	3,529	23.22
Luganda	1.21	362	3.97	1,437	25.64
Afrikaans	0.68	198	4.64	919	13.43
Total	54.41	16,602	4.21	69,928	18.22

Switch points (S*) count alternation points where a token's language tag differs from the preceding token's (Gambäck and Das, 2016). Code-Mixing Index (CMI) follows Das and Gambäck (2014): 0 = monolingual, higher = more balanced mixing.

Note: for Pidgin, which is English-lexified, most tokens read as English, so its CMI/S* understate the mixing a human would perceive; Pidgin audio/transcriptions are taken as-is from the prior release.

Dataset Structure

Each example includes:

• audio: the speech segment, 16 kHz HuggingFace Audio feature (bytes + path)
• language: primary/matrix language of the utterance
• filename: audio clip filename
• transcription: verbatim human transcription (plain text)
• transcription_tagged: same transcription with English spans wrapped in [[EN]]…[[/EN]]
• cmi: per-utterance Code-Mixing Index
• num_switch_points: per-utterance language-alternation count
• duration: utterance length in seconds

Each language is a separate config (subset), selectable in the dataset viewer, each with a single test split. Load e.g. load_dataset("intronhealth/AfriSwitch", "hausa", split="test").

Dataset Creation

Audio was sourced from publicly available YouTube videos and podcasts under permissive licenses; bilingual annotators selected material for the presence of code-switching and produced verbatim transcriptions. For this release the data was processed in two steps:

1. Character-rate filtered — utterances outside each language's 5th–95th percentile of characters-per-second (misaligned audio/text) were removed.
2. Code-mixing metrics added — per-utterance CMI and switch points were computed for each utterance.

The released dataset contains transcriptions and processed audio segments only, with no links back to original source media. No annotator demographic information is included.

Licensing Information

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