Which language pairs does Maraba support for code-switching?

Hausa-English, Yoruba-English, Igbo-English, Pidgin-English, and combinations thereof (e.g. Hausa-Yoruba-English in the same call). All five languages share one model, so any pairing is supported by default.

Code-switching · Production-grade

"Zan zo gobe, but can you confirm the price?"

Q: How does Maraba handle code-switching differently?

Orinode STT (the engine Maraba uses) tags every word independently. A Hausa-English call gets a per-token sequence like ['ha','ha','ha','en','en','en'] with confidence scores. Downstream tasks — TTS response, intent extraction, summary — read the right language for the right span.

The way Nigerians actually speak. Maraba is the only commercial voice AI that handles mid-sentence Hausa-English, Yoruba-English and Igbo-English code-switching — with per-token language labels and no awkward pauses.

Request beta Read the benchmark → See Orinode STT →

Why code-switching is the hardest problem in Nigerian voice AI

In Lagos boardrooms, Kano markets, and Enugu pharmacies, no real conversation stays in one language for long. Code-switching isn't a bug — it's the dominant mode of Nigerian business communication. Yet every off-the-shelf STT model treats it as noise.

🧠

One-language-per-utterance assumption

Whisper, Google Cloud STT, Wit.ai — all commit to a single language label for the whole audio file. They have no way to express "the first three words were Hausa, the rest English."

🔇

Phonology collisions

English vowels are forced into Hausa phonotactics, or vice versa — producing nonsense transcripts. The downstream LLM never sees the actual words.

⏸️

Awkward voice-agent pauses

When the language switches, the agent's TTS pipeline often hesitates — the dead air signals "this AI doesn't understand me" and callers hang up.

📊

Analytics fail silently

Intent extraction, sentiment analysis, keyword search — all read the broken transcript and produce confidently wrong results. You don't even know it's broken.

How Maraba handles it

Per-token language labels, dual-vocab decoder, voice-pair switching. One model, five Nigerian languages, any combination.

1 · Per-token language tagging

Orinode STT's decoder emits a (token, language, confidence) triple at every step. The labels are part of the output, not metadata bolted on after.

"Zan"  ha 0.98
"zo"   ha 0.98
"gobe" ha 0.97
","    —
"but"  en 0.99
"can"  en 0.99
"you"  en 0.99
"confirm" en 0.99
"the"  en 0.99
"price?" en 0.99

2 · Voice-pair switching in TTS

Maraba's response pipeline reads the per-token labels and selects the matching voice variant. The output audio switches phonology cleanly mid-sentence — no acoustic glitches.

# Mixed-language response synthesis
text = "Sannu, your delivery is on the way."
#       ^^^^^                              ^
#         ha                              en-NG

audio = orinode_tts.synthesise(
    text,
    voice_pair=("ha-aisha", "en-ng-tola"),
    detect_per_token=True,
)

Public benchmark — code-switch WER

500-utterance Hausa-English code-switch test set, transcribed by 3 native speakers. WER computed against the consensus reference.

System	WER (overall)	WER (Hausa tokens)	WER (EN tokens)	Per-token lang acc.
Whisper-large-v3	41.7%	58.9%	22.3%	—
Google Cloud STT (ha)	52.4%	46.1%	61.0%	—
Meta MMS-1B-all	34.2%	39.5%	27.8%	—
Orinode STT	12.6%	13.4%	11.8%	97.2%

Full benchmark write-up across all three Nigerian language pairs (Hausa-EN, Yoruba-EN, Igbo-EN) is in our trilingual code-switching benchmark post. Eval notebook + reference transcripts coming on GitHub. Raw scores at /benchmarks.json.

Supported language pairs

All combinations of the five core languages are handled by the same model — no separate fine-tunes per pair.

ha ↔ enHausa-English

yo ↔ enYoruba-English

ig ↔ enIgbo-English

pcm ↔ enPidgin-English

ha ↔ yo ↔ enThree-language mix

Frequently asked

What is code-switching in voice AI?+

Code-switching is when a speaker alternates between two or more languages within a single conversation — often mid-sentence. In Nigeria, this is the default mode of business communication.

Why does code-switching break most STT systems?+

Most speech-to-text models commit to one language label per utterance. When the speaker switches mid-sentence, the model either transcribes the non-dominant words as gibberish or forces them into the wrong language's phonology.

How does Maraba handle code-switching differently?+

Orinode STT tags every word independently. A Hausa-English call gets a per-token sequence like ['ha','ha','ha','en','en','en'] with confidence scores. Downstream tasks read the right language for the right span.

Which language pairs does Maraba support?+

Hausa-English, Yoruba-English, Igbo-English, Pidgin-English, and any three-way combination. All five languages share one model.

Can I access code-switch STT via API?+

Yes — see /products/orinode-stt/. Pass detect_code_switch: true in your request and you'll receive per-token language labels.