SpeechT5

Why we included this project

Microsoft's SpeechT5 repository is a useful place to see how a unified speech-text pre-training idea grew into a whole family of models. The core SpeechT5 encoder-decoder learns from both speech and text, and the codebase adds implementations for end-to-end speech recognition, speech-to-speech and speech translation, text-to-speech, and speech language models such as WavLLM. Each model lives in its own folder with training and inference scripts, so you can reproduce published results or rework the recipes for your own data. If you want to study how modern speech models are built rather than plug in a single off-the-shelf recognizer, this is a practical reference.