Landscape and feasibility
Which bitext, dictionaries, and speech already exist? Who collected them? An honest no is a deliverable.
Consultancy · research · language technology
Tokyma
A consultancy for machine translation and language technology, with a focus on low-resource and Global South languages. We advise on data, models, and field practice. We also edit Malinali, a small offline translation app — one outcome of the same work, not the whole of it.
The industry is moving. The hard languages are still hard.
In a few years, cloud translation and large language models have opened pairs that sat locked for decades. That is real progress. It is also easy to confuse a new dropdown, a demo, or a device SKU with a language that communities can actually use — offline, in confidence, in the orthography they write, in the register they speak.
Tokyma sits with organisations that cannot wait for a vendor catalogue, and with researchers who already know the catalogue is not the same as a corpus. Consulting first: what exists, what is licensed, what is missing, and which tools are the right ones for the job.
How we work
The brief is usually smaller than a platform and larger than a model card.
Corpora and evaluation
Alignment, register, leakage, hold-outs that are not shuffled liturgy. Scores that mean something on the street as well as on a test set.
Models into the field
Fine-tunes, packaging, and the unglamorous path from a checkpoint to a device people will actually carry. When an app is the right surface, we edit Malinali.
Speech as evidence
Oral languages do not become text because a paper needed a table. Collection and transcription have to work where the recording was made.
Neighbours in the field
Low-resource language technology is not empty. Much of the serious work is already being done by communities that do not appear in consumer catalogues.
Masakhane NLP
A grassroots research community for African languages, by Africans. Participatory MT, NER, QA, speech. masakhane.io
CLEAR Global / TWB
Humanitarian language services and language-use data when a crisis does not happen in English. clearglobal.org · language data by country
SIL Global
Orthography, literacy, and speech resources in languages most benchmarks still skip. sil.org
Notes
Short essays on an industry that is changing quickly — coverage, corpora, and problems that were locked for years.
A language that is not in the catalogue is usually the starting point.
We work with research labs, NGOs, and language communities on feasibility, corpora, and tooling. The first reply is a reading of the landscape.