Logo Lanfrica

benardabuto081/sauti-labs

Domaine:

natural language processing

Type de record:

softwareproject
Créateur:
ben
Hôte:
Building Africa's Speech and Language Intelligence Infrastructure # Sauti Labs **Building Africa's Speech and Language Intelligence Infrastructure** Sauti Labs is a research and infrastructure company building the foundational technologies that allow machines to understand, generate, translate, and communicate naturally across African languages. This is not an application. This is infrastructure - the foundation upon which speech and language products across Africa will be built for decades. ## Strategy **Multilingual by destination. Language-first by execution.** Rather than developing many languages shallowly in parallel, Sauti Labs takes one language to genuine depth - data discovery, corpus development, speech acquisition, annotation, benchmarking, and a working baseline model - before moving to the next. **Dholuo is Language 1.** Kiswahili's existing real corpus (110 text documents, 100 speech clips) is preserved as a valuable parallel asset and the foundation for Language 2. ## What's Actually Built (not aspirational) - **8 core frameworks**, each at full maturity: JSON Schema, a central registry, a tested SDK, and CI enforcement on every push. `language-program-framework`, `dataset-framework`, `corpus-registry`, `speech-registry`, `speaker-registry`, `model-registry`, `benchmark-framework`, `annotation-framework`, `api-framework`. - **Real, licensed language data** for two languages, with full provenance tracked through the registries above. - **A split licensing architecture** (Apache 2.0 for infrastructure, preserved original licenses for third-party data) - see `LICENSE`, `NOTICE`, and `languages/LICENSE-DATA.md`. - **An active Dholuo Resource Census** investigating the real Dholuo data ecosystem across academic corpora, open-licensed audio, and broadcast archives - see `docs/Enhancements.md` for findings as they emerge. ## Architecture Framework-first per `docs/architecture/ADR-0001.md`. Privacy-by-architecture for speaker/speech data per `docs/architecture/ADR-0002.md`. Split licensing per `docs/architectur …