Orinode STT · v2.3

Nigerian speech-to-text
that actually understands Hausa.

A single API for Hausa, Igbo, Yoruba, Nigerian English and Pidgin — with diacritic-safe transcripts, real-time language detection, and code-switch handling. Built in Lagos, tuned on telephony audio.

Request beta API key Public benchmarks (JSON) →
11.4%
Hausa WER (vs Whisper 47.2%)
<800ms
Language detection latency
5
Nigerian languages
₦0.50
Per second of audio

What it does

Orinode STT was built for the way Nigerians actually talk — fast, mid-sentence code-switching, telephony bandwidth, real-world noise. Not a thin wrapper over Whisper.

🎙️
5 Nigerian languages
Hausa (ha), Igbo (ig), Yoruba (yo), Nigerian English (en-NG), and Nigerian Pidgin (pcm) — one endpoint, one API key.
🔀
Code-switch handling
"Zan zo gobe, but can you confirm the price?" — Orinode STT keeps the language label per token, not per utterance.
✍️
Diacritic-safe
ƙ, ɗ, ɓ, ts, ƴ (Hausa); ị, ụ, ọ (Igbo); è, é, ẹ, ọ, ṣ (Yoruba) — preserved as UTF-8 in every response.
Streaming + batch
WebSocket for sub-second partial transcripts. HTTP POST for batch jobs and recordings up to 4 hours.
📞
Telephony-grade VAD
Trained on 8 kHz GSM audio with realistic Nigerian network conditions. Handles MTN, Airtel, Glo, 9mobile.
🔐
NDPR-aligned
Audio processed in-region. PII tokens redactable on request. Audit log for every call.

Supported languages

All languages return ISO 639-1 / 639-3 codes per token plus a confidence score.

haHausa
igIgbo
yoYoruba
en-NGNigerian English
pcmNigerian Pidgin

Accuracy — public benchmarks

Reproducible WER on public eval sets. Full methodology + 1,500-utterance code-switch benchmark in our trilingual benchmark post.

LanguageTest setWhisper-large-v3Orinode STTΔ
HausaCommon Voice 17 (ha)47.2%11.4%−35.8
YorubaCommon Voice 17 (yo)38.6%14.8%−23.8
IgboCommon Voice 17 (ig)52.1%17.9%−34.2
Nigerian EnglishNaijaVoices (held-out)18.4%9.2%−9.2
Code-switch (Hausa↔EN)Internal 500-utt eval41.7%12.6%−29.1

Numbers are word error rate (lower is better). Eval methodology + raw outputs published at /benchmarks.json.

Code samples

Three ways to call the API. Pick whichever fits your stack.

PYTHON

# Transcribe a Hausa audio file
import requests

resp = requests.post(
    "https://maraba.ai/api/v1/stt/",
    headers={"X-API-Key": "sk_live_..."},
    json={
        "audio_url": "https://cdn.example.com/call.wav",
        "language_hint": "ha",
        "preserve_diacritics": True,
        "detect_code_switch": True,
    },
)
print(resp.json()["text"])
# "Sannu, ina son ƙwayoyin ciwon kai"

CURL

curl https://maraba.ai/api/v1/stt/ \
  -H "X-API-Key: sk_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "audio_url": "https://cdn.example.com/call.wav",
    "language_hint": "auto",
    "preserve_diacritics": true
  }'

JAVASCRIPT (NODE)

const resp = await fetch("https://maraba.ai/api/v1/stt/", {
  method: "POST",
  headers: { "X-API-Key": process.env.ORINODE_KEY },
  body: JSON.stringify({
    audio_url: url, language_hint: "auto",
  }),
});

Pricing

Pay only for audio processed. No subscriptions, no per-seat fees. Maraba (the call-center product) bundles unlimited STT for its plans.

Per-second
₦0.50 / second of audio. ₦25 minimum per request.
Volume tier
Above 100,000 sec/month: rate drops to ₦0.38/sec.
Beta
Free during private beta — up to 50 hours of audio while we onboard partners.

Frequently asked

Which Nigerian languages does Orinode STT support?+
Hausa (ha), Igbo (ig), Yoruba (yo), Nigerian English (en-NG), and Nigerian Pidgin (pcm). All five share one endpoint and one API key — language detection is automatic per utterance.
How accurate is Orinode STT on Hausa compared to Whisper?+
11.4% WER vs Whisper-large-v3's 47.2% on Common Voice 17 Hausa. The gap widens on telephony audio and code-switched speech.
Does Orinode STT preserve Hausa, Igbo, and Yoruba diacritics?+
Yes. We never .lower() non-English transcripts. ƙ, ɗ, ɓ (Hausa), ị, ụ, ọ (Igbo), è, é, ẹ, ọ, ṣ (Yoruba) are returned as UTF-8.
Can I stream audio for real-time transcription?+
Yes — WebSocket at wss://maraba.ai/api/v1/stt/stream/. Partial transcripts every ~200 ms.
How is Orinode STT different from Maraba?+
Orinode STT is the underlying speech-recognition model. Maraba is the end-to-end AI call-center product that uses Orinode STT, Orinode TTS, and the in-house Maraba LLM. Developers integrating speech-to-text directly use Orinode STT. Businesses that want a hosted receptionist use Maraba.
Is the model open-source?+
Orinode-STT-Small (a 250M-parameter variant) is scheduled for Apache 2.0 release on Hugging Face by Q3 2026. The production model used for Maraba's telephony stays closed.

Start with 5 Nigerian languages on day one.

Request a beta API key — we're onboarding ten partners this quarter.

Request beta API key