NVIDIA-NeMo/NeMo

Provides modular PyTorch-native components for ASR, TTS, and speech LLMs with streaming-capable models like Nemotron-Speech and MagpieTTS supporting multilingual inference. Built on composable neural modules that enable efficient fine-tuning and deployment while integrating with NVIDIA's NIM (NVIDIA Inference Microservices) for production serving. Supports variable latency-accuracy tradeoffs through single checkpoints, enabling researchers to optimize for real-time or high-accuracy requirements.

16,894 stars. Actively maintained with 56 commits in the last 30 days.