01 · Speech recognition
AI research & robotics lab
Neriqlabs builds speech, language and perception models, and the robots and sensors that use them. We start where the world's training data runs out: Ghanaian languages, African environments, and machines that have to work outside the lab.
Why we exist
Today'sAIlearnedfromathinsliceoftheworld:adozenlanguages,afewcities,tidylabs.Everythingoutsidethatslicewasskipped.Webuildthemodels,machinesanddatathathear,seeandmoveintherestofit.
Neriqlabs is a member of the NVIDIA Inception program, which supports startups building on accelerated computing.
The lab
Language AI · the gap
The global AI stack works in about twelve languages. Seven thousand are spoken on Earth.
We went through the documentation of sixteen commercial speech vendors. Not one lists Akan, Twi, Ewe, Ga or Dagbani. So a Ghanaian bank that wants to answer calls by voice has two options: an English-only phone menu many of its customers cannot use, or more human agents. The missing piece is data, and then models trained on it.
Live call · Twi and English
Mepawokyɛw,checkmybalancenomame.
Please check my balance for me.
What we have built
The stack
Three models, one call. Speech in, meaning in the middle, speech back out.
Speech recognition
It hears Twi on a bad phone line.
Fine-tuned from w2v-BERT 2.0 with a character-level CTC head, trained with 8 kHz phone-codec augmentation so it survives a real call rather than a studio recording. Bilingual Twi and English, and it keeps up when a caller switches between them.
Speech synthesis
It answers in Twi, and switches mid-sentence.
Built on the CosyVoice3 backbone and fine-tuned on restored Asante Twi speech. As far as our survey of sixteen vendors goes, it is the only released system that moves between Twi and English inside a single sentence, which is how people actually speak.
Language model
It understands what the caller wants.
A Qwen3.5-9B fine-tune for Twi customer-service dialogue and tool calling: balance checks, transfers, account problems. Scale alone does not fix this. A 27B model zero-shot scores 9 to 10 out of 10 on English tool calls and falls into repetition loops in Twi.
Not yet published
Measured
Twi, read speech
0.0WER
Unseen BibleTTS split
Ghanaian phone audio
0.0WER
WAXAL Akan at 8 kHz
Published baseline
0.0WER
WAXAL-NET Whisper, same data
Twi tool calling
0.00accuracy
0.40 before fine-tuning
Lower word error rate is better. These are development-set figures: a speaker-disjoint held-out test set is still being built, and we will publish it alongside the next release rather than quote a number nobody else can reproduce.
Robotics
Language was our first problem, not our only one. The same gap runs through the physical world: robots and sensors are trained on data from a handful of places and break when they leave them. Our robotics work is perception, state estimation and autonomy for machines that have to work outdoors, at night, in wind and in noise, on hardware cheap enough to deploy.
Pick-and-place · live render
Perception
Cameras, microphone arrays and radio receivers read the same scene. We fuse them into one picture, so a machine is not blind the moment one sensor is.
Estimation
Kalman filtering, multi-target tracking and triangulation from bearings: the maths that turns noisy readings into a position you can act on.
Hardware
Compact compute nodes on single-board computers with AI accelerators, in enclosures we design in CAD and print ourselves. Models run where the sensor is.
Learning
Trajectory labelling, success checks, simulated environments and evaluation for robot learning, with the same paid, audited pipeline we use for speech.
How it works
Engine 01 · Models
We design speech, language and perception models for production: small enough to run where customers and sensors are, fast enough for real conversation and real-time tracking, and built for the languages, accents and environments the largest models miss.
Engine 02 · Data
Low-resource does not only mean language. It means any data the industry never bothered to collect: a dialect, a dialling tone, a factory floor, a robot arm in a workshop in Accra. We design the collection pipelines, pay the people who produce the data, and deliver it licensed, cleaned and evaluation-graded.
Coverage
Twi first, because it is the fastest route to proof: thirty million speakers and no vendor serving them. Ga, Ewe, Dagbani and Hausa are next, and the method travels further than West Africa.
Asante Twi
~30M speakers · models shipped
Ga
~1M speakers · next
Ewe
~5M speakers · next
Dagbani
~3M speakers · next
Hausa
~50M speakers · next
Data partnerships
We run collection and annotation in markets most vendors do not reach, with contributors paid per task and the licensing settled before a single file is recorded. Speech and language is where we started, not where we stop: if the data you need does not exist yet, that is the work.
Research groups working on African languages, robotics teams that need data from new environments, and operators in Ghana: all are the right conversation.
