NeMo-Speech.cpp is NVIDIA's official native C++ runtime for local speech inference, providing day-0 support for models from the NVIDIA NeMo Speech project by running them through ggml, the tensor library that also powers llama.cpp and whisper.cpp. The project positions itself as a single lightweight binary that replaces a Python/PyTorch stack for speech workloads: the nemo-speech CLI ships with indexed defaults for streaming and offline ASR (Nemotron 3.5 ASR Streaming 0.6B, Parakeet TDT 0.6B v3, Parakeet CTC 1.1B), speaker diarization (Streaming Sortformer 4-speaker v2), text and speech translation (Riva Translate 4B Instruct v2, composed into a full ASR-to-NMT-to-TTS pipeline), speech synthesis (MagpieTTS Multilingual 357M with NeMo NanoCodec), and a full-duplex realtime voicechat model (Nemotron Labs VoiceChat 11B) handling audio, transcripts, and tool calling together. A single nemo-speech transcribe call accepts a WAV file, a microphone stream, or a directory, chaining VAD masking, diarization, punctuation/capitalization, inverse text normalization, and translation into one pass; JSON, SRT, and WebVTT output automatically request word-level timestamps, and diarized JSON attaches a 1-based speaker tag to each word. Live capture is compiled directly into the CLI via miniaudio (CoreAudio/WASAPI/ALSA/PulseAudio, no PortAudio), and directory transcription loads the recognizer once and batches concurrent utterances on the GPU. The installer downloads a verified native release for the detected platform and backend, or falls back to a source build with CPU, CUDA, Metal, and Vulkan presets. nemo-speech serve starts a local HTTP server with OpenAI-compatible transcription/speech endpoints plus a realtime WebSocket API and a web playground, a separate riva_server binary exposes Riva-compatible gRPC interfaces, and releases bundle a native C SDK with an exported CMake package so an app can link only the capability it needs. Model downloads are verified against a pinned size and SHA-256, and a doctor subcommand checks the local environment. NVIDIA-authored code is released under Apache License 2.0; bundled third-party components carry their own listed licenses.