Qoru x Lingua Labs: bringing voice-first Igbo NLP to production
We're partnering with Lingua Labs to co-develop an open speech toolkit for underserved West African languages — starting with end-to-end Igbo ASR and TTS.
When we set out to make language models that actually work for underserved communities, we knew we couldn't do it alone. That's why today we're announcing a partnership with Lingua Labs, a Lagos-based startup building real-time speech technology for African languages.
Together we're launching an open toolkit for voice-first applications in West African languages — starting with end-to-end Igbo automatic speech recognition (ASR) and text-to-speech (TTS) that you can run locally.
Why voice matters
Text-based models get all the headlines, but in many communities voice is the primary interface to technology. In Nigeria, 70% of adults speak English as a second language, and hundreds of millions more are far more comfortable using their native tongue. A model that only reads and writes Igbo text misses the point if most people want to just speak.
Lingua Labs has spent the last two years collecting and labeling a 1,200-hour Igbo speech corpus — painstakingly transcribed, time-aligned, and validated by native speakers across five states. What they had was world-class data. What they needed was a team that understood how to turn data into models that generalize. That's where Qoru came in.
How we split the work
| Partner | Expertise | Contribution |
|---|---|---|
| Lingua Labs | Speech data, acoustic domain, UX | Corpus, phonetics, evaluation, product design |
| Qoru | Multilingual modeling, open releases | Model architecture, training infrastructure |
Lingua Labs led data collection, phonetic annotation, and user testing. Qoru contributed the modeling architecture — a conformer-based ASR stack and a flow-matching TTS pipeline — along with the compute and engineering bandwidth to train and ship the first checkpoints.
The first release
Today we're publishing:
lingua-igbo-asr-v1— a streaming conformer ASR model, 48ms word error rate on our held-out test set, runs on a mid-range Android phone.lingua-igbo-tts-v1— a zero-shot voice model that clones a speaker from a 10-second sample, with a 24kHz neural vocoder.
Both are Apache 2.0 and available on huggingface.co/qoru. The training scripts, the corpus, and the full evaluation suite are open too — we want other teams to build on this, not rebuild it.
from transformers import AutoModel, AutoProcessor
processor = AutoProcessor.from_pretrained("qoru/lingua-igbo-asr-v1")
model = AutoModel.from_pretrained("qoru/lingua-igbo-asr-v1")
waveform = load_audio("my-recording.wav")
prediction = processor.decode(model(waveform).argmax(dim=-1))What's next
This is our first collaboration, and we're treating it as a template. Over the next six months we plan to:
- Add Yoruba and Hausa to the toolkit (data collection starts next month).
- Ship a mobile SDK so app builders can embed Igbo voice in minutes.
- Open a joint grants program for community groups building Igbo-language applications.
If you're a researcher, builder, or funder working in voice for underrepresented languages, we'd love to hear from you. Reach out to Daniel at hi@qoru.ai or to Chinedu at Lingua Labs (chinedu@lingualabs.africa).
Two teams, one mission: language technology that works for everyone.