Low-Resource Is Not a Property of the Language

There is a phrase that appears constantly in NLP literature: "low-resource language," meaning the lack of "training data that is required to build accurate state-of-the-art NLP systems."¹ (See also: Besacier et al., 2014)² It is applied to Amharic, for example, a language with over 30 million native speakers generating speech every single day. The phrase describes an attention and investment gap, not a property of the language. Ethiopian languages are low-resource because they have been low-priority.

Working with a low-resource language often means optimizing for scarcity: small models, clever tricks, transfer learning etc. Recognizing the scarcity as a result of inattention means doing what we did. We build the data infrastructure, build the tokenizers, build the collection platforms, build the QC systems, and train models that work. The approach is not novel.

One reason it has not been done for a long time is structural. The organizations with the compute to build speech models for 80+ Ethiopian languages do not have the linguistic knowledge, the local infrastructure, or the incentive to do it well. The resulting Amharic models check a box on a coverage chart³. They break on real-world audio because they were validated on read speech by a narrow set of speakers. Meanwhile, the people with the linguistic knowledge have not had the engineering infrastructure or the funding to build at production scale.

Phonetico exists at the intersection of both. We will be publishing more about our work, including one on multi-task CTC training for Amharic ASR, and releasing a portion of our datasets publicly as they reach maturity. The first will be a Tigrinya speech dataset.

There is more to come.