Emperor-Trillion/yoruba-fastspeech2

FastSpeech2 for short-utterance Yoruba speech synthesis # yoruba-fastspeech2 **FastSpeech2 for short-utterance Yoruba speech synthesis.** Existing Yoruba TTS systems are trained on full sentences and fail on isolated phonemes, short words, and brief phrases — the output types most needed for language learning, dictionary, and educational tools. This repository provides a dedicated FastSpeech2 model trained from scratch on 2,466 short Yoruba utterances using MFA-derived phoneme alignments, complementing sentence-trained models for sub-sentence inputs. Paper: _FastSpeech2 for Short-Utterance Yoruba Speech Synthesis: A Low-Resource Pipeline_ — Sunday Emmanuel Sanni, 2026. --- ## Quickstart (inference in 5 steps) ```bash # 1. Clone this repo and the FastSpeech2 base git clone github.com git clone github.com cd FastSpeech2 # 2. Apply the Yoruba patches patch -p1 **Tone not encoded in phoneme tokens.** The preprocessing pipeline stripped `_H`/`_L` tone markers before training, so `bàbá` (low-high) and `baba` (mid-mid) produce identical input sequences. The model cannot distinguish them at the input level and outputs the statistically dominant tone pattern for each phoneme context. This is a known architectural limitation. Fixing it requires retraining with tone tokens retained — this is the top priority for the next version. --- ## Acknowledgements - ming024/FastSpeech2 — PyTorch implementation used as the base - Niger-Volta LTI / yoruba-voice — G2P vocabulary (6,798 base entries) - Montreal Forced Aligner — phoneme alignment tool - The speaker, Sunday Emmanuel Sanni, for recording 2,466 utterances --- ## Citation ```bibtex @misc{sanni2026yoruba, title = {FastSpeech2 for Short-Utterance Yoruba Speech Synthesis: A Low-Resource Pipeline}, author = {Sanni, Sunday Emmanuel}, year = {2026}, howpublished = {\url{github.com, note = {Code, checkpoints, and demos available onl …