Igbo Speech Recognition
Igbo is the hardest of Orinode's six supported languages — 39.8% WER on Maraba v1 (May 2026). We're transparent about it: low-resource ASR for tonal languages with severe pre-training under-representation is genuinely difficult, and we publish the numbers rather than hide them.
The Igbo language
Igbo is spoken by approximately 30 million people, primarily in southeastern Nigeria. Like Yorùbá, it is a tonal language — but Igbo has additional complexity:
- Two-tone system with downstep — the same syllable can carry high, low, or downstepped-high tone, all phonemically distinct.
- Vowel harmony — verbs alternate between two vowel sets based on a [±expanded pharynx] feature.
- Severe pre-training scarcity — Whisper-large-v3 has no Igbo language token in its vocabulary. We repurposed the rarely-used Lingala (
ln) slot and re-taught the model that signal == Igbo. - Diacritic-heavy orthography — Igbo's standard orthography uses sub-dot diacritics (ọ, ụ, ị) and tone marks; web text is often un-diacriticized, creating tokenizer mismatch.
Architecture notes
Igbo uses the same Speech-LLM stack as the other Maraba v1 languages — Whisper-large-v3 encoder + MLP adapter + Gemma-2-9b decoder + LoRA. The Igbo-specific engineering:
- Token-slot remapping: Whisper's
<|ln|>token is overloaded to mean Igbo. Inference must respect this mapping or the model emits Lingala-flavored output. - Dataset: Mozilla Common Voice 17 Igbo partition + Orinode-curated audio from Igbo radio archives and field recordings.
- Tone-aware loss: standard cross-entropy with label smoothing 0.05.
Performance (May 2026)
| Metric | Value | N |
|---|---|---|
| WER (normalized) | 39.80% | 200 |
| WER (raw, case-sensitive) | 45.11% | 200 |
What we're doing about it
Igbo is the priority for Maraba v2 (target Q3–Q4 2026):
- ~100 hours of newly recorded studio Igbo audio across Onitsha, Enugu, and Owerri dialects.
- Custom Igbo subword vocabulary (replacing the repurposed Lingala slot with a real Igbo token).
- Tone-conditional acoustic modeling layer borrowed from recent African ASR research.
- Target: under 25% WER on Igbo before Maraba v2 ship.
Get the model
Maraba v1 research weights including the Igbo configuration are on huggingface.co/Orinode. For pilot API access: [email protected].