MaryAdelua/yoruba-speech-tools

# Yoruba pronunciation resource and integration research This repository develops a training-ready Yoruba pronunciation resource and a model-independent methodology for improving Yoruba pronunciation in existing multilingual speech-generation systems. ## Research question Can verified Yoruba speech, pronunciation representations, and—when required—explicit tone supervision measurably improve held-out Yoruba pronunciation in an existing multilingual speech model without reducing intelligibility or naturalness? ## Current checkpoint - 120 controlled single-speaker Yoruba recordings collected; - audio quality checked and standardized locally; - Unicode-safe orthographic H/M/L tone annotations generated; - frame-level pYIN F0 tracks and utterance contours extracted; - 17 meaning-relevant contrast groups defined; - utterance-level alignment completed; - all 120 recordings listening-verified by the fluent speaker; - one trailing-noise correction completed (`0205`), with no re-recordings required; - matched ChatGPT Voice and Microsoft Copilot Voice baselines evaluated; - a concise functional-load tone-prompt intervention implemented and tested; - prompt-level tone guidance produced **no aggregate improvement** over the matched ChatGPT baseline: exact transcript accuracy, tone correctness, and naturalness were unchanged, while strict meaning recovery decreased from 33.3% to 22.2% on nine valid pairs. The prompt intervention is retained as a reproducible null result. It shows that prompt wording alone did not repair pronunciation. Future experiments will therefore compare verified data adaptation, a pronunciation-aware frontend, and optional auxiliary supervision in that order. ## Pronunciation resource package The next phase is organized around a portable dataset and method rather than a standalone voice model: - `dataset/` contains the resource specification, annotation guide, split/leakage policy, dataset card, release gates, schema, and metadata-only manifest; - …