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

Dataset Description

AfriSwitch is a 56.35-hour, human-transcribed 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.

• Languages: Amharic, Pidgin, Kinyarwanda, Yoruba, Hausa, Oromo, Igbo, Zulu, French, Shona, Swahili, Tswana, Luganda, Afrikaans (each code-switched with English)
• License: CC BY NC SA 4.0
• Total duration: 56.35 hours
• Total utterances: 17,317
• Total code-switch events: 63,400

Dataset Statistics

Language	Hours	Utterances	Avg. Switch Points (S*)	Total S*	CMI
Amharic	5.00	1,232	4.16	5,130	12.38
Pidgin	5.00	2,003	4.23	8,482	30.28
Kinyarwanda	5.00	1,587	3.20	5,084	13.12
Yoruba	5.00	1,885	3.61	6,805	18.26
Hausa	5.00	1,487	3.34	4,961	11.61
Oromo	5.00	1,464	2.51	3,671	12.44
Igbo	4.99	1,883	3.19	6,003	24.58
Zulu	4.99	1,479	3.94	5,827	23.04
French	4.07	1,158	2.16	2,506	10.08
Shona	3.97	1,184	4.30	5,092	23.05
Swahili	3.93	652	9.18	5,988	24.51
Tswana	2.71	798	2.82	2,247	22.58
Luganda	1.24	374	3.36	1,256	23.92
Afrikaans	0.45	131	2.66	348	9.51
Total	56.35	17,317	3.66	63,400	19.06

Switch points (S*) count alternation points where a token's language tag differs from the preceding token's, following Gambäck and Das (2016).

Code-Mixing Index (CMI) follows Das and Gambäck (2014): a CMI of 0 indicates a fully monolingual utterance, while higher values reflect a more balanced mix of languages within an utterance. Across the languages, mixing intensity varies considerably: Afrikaans, French, and Hausa are closest to monolingual (lowest CMI), while Pidgin and Igbo show the densest, most balanced mixing with English.

Dataset Structure

Each example in the dataset includes:

• audio: the speech segment (VAD-segmented, concatenated up to 30 seconds per utterance to increase the likelihood of capturing a code-switch), as a 16 kHz HuggingFace Audio feature
• filename: the audio clip filename
• transcription: verbatim human transcription
• duration: utterance length in seconds
• language: the primary/matrix language of the utterance

Each language is provided as a separate config (subset), selectable in the dataset viewer.

Dataset Creation

Source Data

Audio was sourced from publicly available YouTube videos and podcasts under permissive licenses. Bilingual annotators on African crowdsourcing platforms selected source material specifically for the presence of code-switching.

The released dataset contains transcriptions and processed audio segments only, with no links back to original source videos or podcasts, preventing direct tracing to original content creators. No annotator demographic information is included in the released benchmark.

Licensing Information

This dataset is released under the Attribution-NonCommercial-ShareAlike 4.0 (CC BY NC SA 4.0 ) license.