| --- |
| license: cc-by-nc-sa-4.0 |
| task_categories: |
| - automatic-speech-recognition |
| language: |
| - pcm |
| - yo |
| - ha |
| - sw |
| - en |
| tags: |
| - code-switching |
| - african-languages |
| - speech |
| - telco |
| pretty_name: SwitchBoard Tier B — African code-switched speech |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: data/*.wav |
| --- |
| |
| # SwitchBoard Tier B — African code-switched speech |
|
|
| 87 consented utterances of **intra-sentential code-switching** — Nigerian Pidgin, |
| Yorùbá, Hausa and Kiswahili each mixed with English inside a single sentence — |
| recorded from 8 bilingual volunteers at the Deep Learning Indaba 2026, Lagos. |
|
|
| Collected for the MLC (Africa) × Intron Agentic Voice AI Challenge as an |
| evaluation set for telco/fintech voice agents. **8.75 minutes total.** |
|
|
| ## What this is for |
|
|
| Measuring whether a speech model survives the *switch points* — and whether the |
| errors it makes are the ones that change a transaction. Every clip carries |
| per-token language tags, gold intent and slot labels, so you can compute PIER |
| (error rate restricted to embedded-language tokens), Entity-WER and Numeric-WER, |
| not just WER. |
|
|
| ## Composition |
|
|
| | pair | clips | speakers | |
| |---|---|---| |
| | pcm-eng (Nigerian Pidgin × English) | 30 | 2 | |
| | yor-eng (Yorùbá × English) | 28 | 2 | |
| | swa-eng (Kiswahili × English) | 20 | 2 | |
| | hau-eng (Hausa × English) | 9 | 2 | |
|
|
| 77 clips Nigeria, 10 Kenya · 49 female, 38 male · 16 kHz mono WAV · 3.1–11.2 s |
| (median 5.7 s) · 2 domains (telco 57, fintech 30) · 2 noise conditions (quiet 49, |
| ambient 38). |
|
|
| ## Fields |
|
|
| `file_name`, `transcription`, `speaker_id` (pseudonymous `SPK-xxxxxxxx`), |
| `prompt_id`, `pair`, `language`, `duration_s`, `domain`, `intent`, `switch_type`, |
| `accent`, `country`, `gender`, `age_range`, `noise_condition`, `device_type`, |
| `slots` (JSON), `entities` (JSON), `lang_tags` (space-separated, one tag per |
| whitespace token of `transcription`). |
|
|
| `lang_tags` aligns 1:1 with `transcription.split()`. That alignment is enforced |
| by the build script; a mismatch would shift the point-of-interest mask and make |
| PIER measure the wrong tokens. |
|
|
| ## Limitations — stated as defects, not caveats |
|
|
| 1. **hau-eng is effectively one speaker.** One volunteer completed 8 of 8 prompts; |
| the second stopped after 1. Any Hausa-specific number rests on 9 clips. |
| 2. **No device diversity.** All 87 clips are one USB headset. Real telco audio is |
| 8 kHz narrowband over a handset; nothing here measures that. |
| 3. **Read speech, not conversation.** Volunteers read prompts. No disfluency, |
| barge-in, turn-taking or spontaneous repair. |
| 4. **8 speakers, 2 per pair.** Accent, device and noise floor are confounded with |
| speaker identity. Block your bootstrap on speaker; it will not be narrow. |
| 5. **Geographic concentration.** 77 of 87 clips from Nigeria. |
| 6. **Pidgin language tags are the least reproducible.** Nigerian Pidgin is |
| English-lexified, so the matrix/embedded boundary is a semantic judgement over |
| orthographically English words. |
|
|
| **Too small to train on, and too small to rank systems with.** Use it to probe |
| behaviour at switch points, not to declare a winner. |
|
|
| ## Consent and privacy |
|
|
| Every speaker gave explicit affirmative consent to this exact text, stored |
| verbatim with a version and timestamp alongside their record: |
|
|
| > *"I have read the information about this research recording. I agree to my voice |
| > recordings and their transcripts being published under CC BY-NC-SA 4.0 for |
| > research on African speech recognition. I understand no personal information is |
| > spoken, and that I may have my recordings deleted at any time by contacting the |
| > researcher."* |
|
|
| Prompts contain **no real personal data** — the names and phone numbers spoken |
| were written for the script, and no volunteer spoke their own name, number or any |
| account detail. Speaker identifiers are pseudonymous; the name↔id mapping is held |
| offline solely so a deletion request can be honoured, and has never been |
| published. |
|
|
| **One correction, found by audit.** We previously described the spoken names as |
| "invented". One is not: prompt `yor-010` uses *"Adebayo Ogunlesi"*, the name of a |
| real and well-known Nigerian businessman. Both parts are ordinary Yorùbá names and |
| the collision was accidental — the prompt is a routine "my name is X, my number is |
| Y" customer-service line and says nothing about that person — but "invented" was |
| the wrong word. The audio is already recorded and consented, so it has not been |
| altered; the name is flagged here and a re-record should replace it. |
|
|
| ## Withdrawal |
|
|
| Any speaker may have their clips removed. Contact the maintainer; clips are |
| deleted and the release id is incremented. |
|
|
| ## Citation |
|
|
| ``` |
| @misc{switchboard_tierb_2026, |
| title = {SwitchBoard Tier B: African code-switched speech for voice-agent evaluation}, |
| author = {Daudu, Moses}, |
| year = {2026}, |
| note = {Deep Learning Indaba 2026, Lagos. CC BY-NC-SA 4.0.} |
| } |
| ``` |
|
|
