AfriVoices-KE is a large-scale multilingual speech dataset comprising approximately 3,000 hours of audio across five Kenyan languages: Dholuo, Kikuyu, Kalenjin, Maasai, and Somali. The dataset includes 750 hours of scripted speech and 2,250 hours of spontaneous speech, collected from 4,777 native speakers across diverse regions and demographics. This work addresses the critical underrepresentation of African languages in speech technology by providing a high-quality, linguistically diverse resource. Data collection followed a dual methodology: scripted recordings drew from compiled text corpora, translations, and domain-specific generated sentences spanning eleven domains relevant to the Kenyan context, while unscripted speech was elicited through textual and image prompts to capture natural linguistic variation and dialectal nuances. A customized mobile application enabled contributors to record using smartphones. Quality assurance operated at multiple layers, encompassing automated signal-to-noise ratio validation prior to recording and human review for content accuracy. Though the project encountered challenges common to low-resource settings, including unreliable infrastructure, device compatibility issues, and community trust barriers, these were mitigated through local mobilizers, stakeholder partnerships, and adaptive training protocols. AfriVoices-KE provides a foundational resource for developing inclusive automatic speech recognition and text-to-speech systems, while advancing the digital preservation of Kenya’s linguistic heritage.

Sign language corpora face a structural tension between open-access requirements and the irreducible biometric identity embedded in visual, gestural data. While 3D pose estimation enables signer-agnostic abstraction, the representational adequacy of pose-based modeling for preserving linguistic structure remains underexplored. This paper introduces a perceptually-grounded kinematic modeling framework that formalizes 3D landmark sequences as an intermediate linguistic representation and validates their adequacy through avatar-mediated synthesis and large-scale human evaluation. Using 30370 gloss-level Kenyan Sign Language (KSL) segments derived from the AI4KSL corpus, we construct normalized 3D motion trajectories via MediaPipe Holistic. These trajectories are retargeted to parameterized avatars through a constrained kinematic mapping that preserves non-manual marker geometry and articulatory timing. We define a dual evaluation paradigm combining geometric fidelity metrics (PCK=92.7%, OKS=0.88, PCP=91.5%, PDJ>85.3%) with perceptual constructs measured across a statistically powered Deaf participant cohort (N=384). Results demonstrate a strong predictive relationship between structural joint precision and perceived gesture clarity (r=0.76, p<.01), suggesting that linguistic adequacy is partially recoverable from normalized kinematic structure. Furthermore, representational diversity in avatar instantiation significantly increases perceived inclusivity without degrading intelligibility. These findings establish pose-based motion abstraction not merely as an anonymization technique but as a viable corpus-level modeling layer for ethically sustainable language in motion.