lingvo

Why we included this project

Lingvo is a TensorFlow framework for sequence models, the kind Google's research teams used to build and publish a long string of their speech recognition, machine translation, and language modeling work. It is not an application you install and run; it is a set of modular layers, attention mechanisms, and ready-to-run experiment configs that cover ASR, translation, language modeling, and even 3D object detection. The design centers on reproducibility and distributed training across GPUs and TPUs, which suits research groups that want to modify a model architecture and then scale the result. The tradeoff is that it stays close to TensorFlow and makes no promises about backwards compatibility between releases, so it rewards engineers who already work in that ecosystem and want a proven reference implementation rather than a supported product.