sidneyowallah/dholuo-tts

Dholuo POS-Aware TTS Pipeline # Dholuo POS-Aware TTS Pipeline 🇰🇪 This repository contains a complete pipeline for building a high-fidelity, natural-sounding Text-to-Speech (TTS) system for the **Dholuo** language. Unlike standard multilingual TTS, this system uses a **Part-of-Speech (POS) aware G2P (Grapheme-to-Phoneme)** approach to solve the problem of homograph disambiguation and tonal accuracy. ## 🎯 Try the Demo **🚀 Live Demo** Experience the Dholuo TTS system in action with our interactive demo featuring male and female voice models. --- ## 🌟 Features - **Grammatical Disambiguation** Uses a fine-tuned **AfroXLMR** model to identify POS tags, allowing the system to distinguish between homographs (e.g., _dhok_ as "mouth" vs. "cows"). - **Tonal Injection** Automatically injects tone markers (`˥` for high, `˩` for low) into the phonetic transcription based on grammatical categories (Nouns vs. Verbs). - **Custom Dholuo G2P** A specialized phonemizer that handles Dholuo-specific digraphs (`ny`, `ng'`, `th`, `dh`) and ATR vowel harmony. - **Gender-Specific Voice Models** Separate training pipelines for male and female voices, with automatic dataset splitting by speaker gender. - **Modular Architecture** Separated concerns with dedicated modules: `phonemizer.py` for G2P, `tagger.py` for POS tagging, and `utils.py` for shared utilities. - **Comprehensive Testing** Full test suite covering unit tests, integration tests, and end-to-end pipeline validation with gender-specific model testing. - **End-to-End VITS** Training on the state-of-the-art **VITS** (Variational Inference with adversarial learning for end-to-end Text-to-Speech) architecture. - **Interactive Demo** Next.js web app (in `web/`) on Vercel, backed by the FastAPI inference Space. - **Production-Ready API** FastAPI backend with Redis caching and Nginx reverse proxy. - **Dockerized Deployment** Simple deployment using Docker and Docker Compose. - **Hugging Face Integration** Automatic model downloading from the Hug …