NeMo-Speech.cpp
A lightweight native C++ inference runtime for NVIDIA's Nemotron speech model family, covering streaming speech recognition, speaker diarization, speech translation and text-to-speech, powered by ggml with broad hardware support.
NeMo-Speech.cpp is a lightweight, native C++ runtime for running NVIDIA Nemotron speech models locally, built on ggml for broad hardware support. It is positioned as NVIDIA's own solution for local speech inference, with day-0 support for new models from the NeMo Speech project.
It covers several applications: streaming speech recognition with the Nemotron ASR and Parakeet model families, speaker diarization with Sortformer and Nemotron Diarization, speech and text translation through Riva Translate combined into a full speech-to-speech pipeline, speech synthesis with MagpieTTS and NanoCodec, and a full-duplex voice chat mode with realtime audio, transcripts and tool calling. It also includes Silero voice-activity detection, punctuation and capitalization restoration, endpointing and subtitle generation.
A nemo-speech command-line tool installs through a shell or PowerShell script on Linux, macOS and Windows, preferring a verified native release and falling back to a source build (requiring CMake 3.26 or newer, Ninja, a C++17 compiler and SentencePiece) when no prebuilt artifact is available; the first transcription automatically downloads and verifies the default model from Hugging Face.