tsionshamsu/amharic-speech-lab

# amharic-speech-lab ## Overview **amharic-speech-lab** is a research-oriented personal project of mine focused on **automatic speech recognition (ASR)** and **text-to-speech (TTS)** for **Amharic**, a low-resource language that remains underrepresented in modern speech systems. The project currently implements a clean, end-to-end baseline pipeline using Meta’s **Massively Multilingual Speech (MMS)** models: - YouTube audio → WAV preprocessing - Amharic ASR (speech → text) - Amharic TTS (text → speech) The goal of this repository is not only to demonstrate a working pipeline, but to serve as a foundation for future **fine-tuning, evaluation, and voice modeling experiments** for Amharic. ## Motivation As kids, my brothers and I used to treat tools like Google Translate as a game of sorts. We grew up in Ethiopia, so Amharic was our mothertongue and English a close second. When we first started using google translate it was in earnest, and though it worked well in teaching us a few French phrases, we quickly figured out that its abilities were quite poor when it came to Amharic. So we started typing in the oddest Amharic words we would think of and then laughing at the strange, unintentionally funny outputs it would conjure up. It was extremely entertaining, but it also made clear how limited existing systems were when it came to the language. Alongside this, my interest in languages more broadly (shaped in part by studying Latin) and my background as a computer science major naturally pulled me toward machine learning as a way to think about language as a structured, evolving system. Amharic is spoken by tens of millions of people, yet it remains low-resource in many modern speech models. This project is an initial attempt to explore that gap by building and evaluating strong multilingual ASR and TTS baselines, understanding where they break down, and iterating from there. It’s very much a work in progress (there are multiple rough edges and occasional mistake …