Orinode/naija-customer-call-code-switch

Naija Customer-Call Code-Switch Corpus (Orinode-CCS)

Hand-written customer-service sentences with natural code-switching between Nigerian English and three indigenous Nigerian languages — Hausa, Yoruba, and Igbo. Covers 30+ business sectors typical of real customer-service calls in Nigeria.

Released by Orinode under CC-BY 4.0 to support research on multilingual ASR, NLU, and conversational AI for low-resource African languages.

---

Why this dataset exists

Global voice-AI systems trip badly on Nigerian speech. Not because Nigerian languages are obscure — Hausa alone has ~80M native speakers — but because public datasets capturing the way Nigerians actually talk, especially the rapid mid-sentence code-switching between English and indigenous languages, are scarce.

This corpus targets that gap: 15,000 hand-written, naturalistic customer-service sentences a Nigerian would plausibly say to a call-center, broken down by language pair and code-switch ratio.

---

Languages & sizes

Pair	Code	Lines	Format
Hausa ↔ English	ha-en	5,000	One sentence per line
Igbo ↔ English	ig-en	5,000	One sentence per line
Yoruba ↔ English	yo-en	5,000	One sentence per line
Total		15,000

Yoruba sentences use full diacritics (ọ̀ ọ́ ẹ̀ ẹ́ ṣ etc.). Igbo sentences use standard orthography (ọ ụ ị ṅ). Hausa sentences use Latin orthography.

---

Structure (9-section template per file)

Each file follows the same 9-section template so subsets can be cleanly extracted:

Lines	Section	Style
1–500	Gold seed	Natural mixed code-switch, balanced
501–1000	Target-leaning	~70% indigenous language, ~30% English
1001–2000	Balanced	50/50 split
2001–3000	English-leaning	~70% English, ~30% indigenous
3001–4000	Natural mixed	Free mixing
4001–4250	Multi-turn snippets	Sequential lines = same caller
4251–4500	Government services	Passport, FRSC, NIMC, NEPA, customs
4501–4750	Legal / religious / security	Lawyer, mosque, church, NSCDC, embassy
4751–5000	Misc Nigerian specialty	Vet, funeral planner, photography, etc.

The 30 core sectors used in lines 1–4000 include: restaurant, hotel, catering, bakery, event hall, supermarket, online shopping, fashion boutique, phone shop, pharmacy, hospital, dental, lab, salon, spa, gym, bank, microfinance, insurance, telecom, internet, cable TV, petrol/gas, ride-hailing, logistics, airline, mechanic, real estate, home services, education.

---

File format

Plain UTF-8 text. One sentence per line, prefixed with its 1-indexed number:

1. <sentence>
2. <sentence>
...
5000. <sentence>

Sentence durations target 4–25 seconds of spoken Nigerian speech at ~2.5 words/sec (so a 12-word line ≈ 5 sec; a 50-word line ≈ 20 sec). This makes the corpus directly suitable as ASR transcript prompts when paired with audio recordings.

---

Sample sentences

Hausa ↔ English:

Sannu, ina son in yi reservation for tonight, table for two at 7pm.

Igbo ↔ English:

Ndewo, achọrọ m ka m book table for two at Genesis Restaurant Onitsha echi abalị.

Yoruba ↔ English:

Bawo, mo fẹ́ book table for two at Nkoyo Lekki this Saturday evening.

---

Intended use

• Multilingual ASR training (when paired with audio recordings)
• Code-switching detection and language identification
• Customer-service NLU and intent classification
• TTS prompt corpora for Nigerian languages
• Evaluation benchmarks for low-resource African speech systems
• Conversational-AI fine-tuning for Nigerian-market voice assistants

---

Quality notes (honest internal assessment)

Quality is not uniform across the 5,000-line files. We are publishing the full corpus rather than only the high-quality subsets, so users can make their own filtering decisions.

File	Lines 1–3000	Lines 3001–4500	Lines 4501–5000
Hausa-English	Gold quality throughout	Gold quality	Gold quality
Igbo-English	Gold quality	Acceptable; templating in places	Weaker — some Hausa-language leakage, template loops
Yoruba-English	Gold quality	Acceptable	Weaker — ceremony/festival sections repetitive

Recommendation for downstream training: prefer the first 3,000 lines of each file for highest-quality subsets. The middle and later sections (3001–5000) cover broader vocabulary (government, legal, religious, Nigerian specialty) but show more pattern repetition — useful for breadth, less so for quality benchmarks.

We welcome feedback and pull requests at github.com/Orinode-ltd.

---

Methodology

These sentences were authored by hand — no LLM generation, no web scraping, no machine translation of monolingual corpora. Each sentence was written line-by-line by Nigerian-fluent contributors, then reviewed against the section ratio targets.

Code-switching follows natural Nigerian patterns: proper nouns (places, brands, currencies, numbers) stay in English even in target-language-heavy sentences; verbs and connectives follow the section ratio; greetings use the indigenous language (Sannu, Ndewo, Bawo).

---

Limitations

• Text only. No audio. This is a transcript corpus. Audio recording is a planned follow-up.
• Customer-service domain. Not suitable as-is for general conversational AI in other domains (medical, legal, casual chat).
• Lagos / Southwest / Igbo-East / Northern dialects of these languages exist — the corpus reflects the standard written form most callers would use, not regional spoken variation.
• No personal data. All names, phone numbers, account numbers are fictional.

---

Citation

If you use this dataset, please cite:

@misc{orinode_ccs_2026,
  title  = {Naija Customer-Call Code-Switch Corpus},
  author = {{Orinode Ltd}},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/Orinode/naija-customer-call-code-switch}
}

---

License

CC-BY 4.0 — free for commercial and non-commercial use with attribution.

---

Contact

• Research: research@orinode.ai
• Partnerships: partnerships@orinode.ai
• Website: https://orinode.ai
• Org: https://huggingface.co/Orinode

Orinode Ltd · RC 9486856 · Lagos, Nigeria.