SALT Documentation¶

Welcome to the official documentation for the SALT project, part of the Sunbird AI Language Projects.

This documentation covers our datasets, open-source models, and the SALT Python library designed to support speech and language technology for African languages.

Datasets¶

We build and curate high-quality datasets to advance research and application of machine learning for low-resource African languages: - SALT: A multi-way parallel text and speech corpus covering English and six widely spoken languages in Uganda and East Africa. - SALT-31: A context-aware Machine Translation evaluation benchmark covering 31 Ugandan and regional languages. - Urban Noise Uganda 61k: A dataset for urban environmental acoustic monitoring in Uganda.

Models¶

We release highly-optimized models trained for translation, speech recognition, and synthesis: - Sunflower-14B & Sunflower-32B: Our flagship multilingual language models for Ugandan languages and English, including various quantized formats (FP8, W8A8, FP4A16, GGUF). - Whisper Large v3 SALT: An Automatic Speech Recognition (ASR) model fine-tuned on Ugandan languages. - Orpheus 3B TTS Multilingual: A Text-to-Speech (TTS) model supporting voice generation across Ugandan languages. - SunbirdTutor Gemma 4 E2B: A specialized educational model.

📦 SALT Python Package¶

The salt Python package provides helper utilities and pipelines for convenient experimentation, training, and deployment: - Getting Started: Read the Overview & Installation guide. - Developer Guides: Master Data Loading, Data Preprocessing, Model Training, and Evaluation Metrics. - Core Pipelines: See guides for Translation Models, ASR Models, TTS Models, and Speaker Diarization.