NVIDIA/OpenSeq2Seq

Built on TensorFlow, it provides encoder-decoder building blocks optimized for distributed multi-GPU/multi-node training and mixed-precision computation on Volta/Turing GPUs. Supports neural machine translation, automatic speech recognition, speech synthesis, and language modeling through configurable model architectures. Integrates Horovod for scalable distributed training and incorporates beam search decoding with language model re-scoring.

1,560 stars. No commits in the last 6 months.