Magana — Hausa for “speech”, “talk”, “utterance” — paired with Roots: a foundational infrastructure layer we are building intentionally from the ground up for African phonetics, morphosyntactic structures, and accents.
Roots Magana is an ongoing research initiative exploring a speech foundation model designed specifically for the linguistic realities of the continent. We are investigating an acoustic-to-text fine-tuning layer that natively handles code-switching and accented English/French, alongside a non-autoregressive conditional flow-matching TTS model that aims to respect tonal variations and diacritics.
The Roots Magana Engine is the proposed overarching production-grade STT/TTS pipeline — with specialised layers for ASR, TTS, and deployment at the edge, all currently under active investigation.
The proposed overarching production-grade STT/TTS pipeline — orchestration, routing, and quality assurance for every speech interaction. Currently in the design and prototyping phase.
The acoustic-to-text fine-tuning layer we are developing for code-switching and accented English/French — built on African phonetics. Currently exploring data collection and tokenizer design.
A non-autoregressive conditional flow-matching model we are exploring to handle tonal variations and diacritics natively. Early-stage experimentation with flow-matching architectures.
As part of ongoing research, we are evaluating different variations for different production requirements — from high-accuracy cloud processing to low-latency streaming on local telecom IVRs.
A potential full server-side foundation model utilising an expanded morphology-aware tokenizer for high-accuracy cloud processing. We are evaluating its feasibility for enterprise ASR and TTS workloads where precision matters.
An INT8/FP8 quantized execution engine direction we are exploring for low-latency streaming inside local telecom IVRs and offline devices — aiming for real-time speech without the cloud round-trip.
Interested in collaborating on African speech foundation research, exploring the Roots Magana Engine, or discussing edge deployment directions? Let’s talk.