Logo Lanfrica

williampepple1/ibani-tts

Domain:

natural language processing
Creator:
wil
Host:
Ibani Text to Speech # Ibani Text-to-Speech (TTS) Models This repository is dedicated to preparing datasets and training Text-to-Speech (TTS) models for the Ibani language. It is kept separate from the text translation models to prevent Python dependency conflicts between audio libraries and text libraries. ## Data Preparation on Kaggle Since the Ibani audio dataset consists of chapter-level MP3s hosted on Cloudflare, the audio needs to be split into verse-level chunks to train a TTS model. 1. Create a Kaggle Notebook with a GPU (e.g., T4 x2) and Internet enabled. 2. Upload or copy the contents of `kaggle_prep_script.py` into a Kaggle cell. 3. Run the script to automatically: - Download the `ibani_eng.json` text data. - Download the Chapter MP3s from the Cloudflare bucket. - Slice the audio based on silent pauses between verses. - Generate the required `metadata.csv`. ## Requirements The `requirements.txt` is specifically for local audio preparation if you choose to run it locally. For training, a completely different set of TTS libraries (like Coqui-TTS) will be required.