nemotron-speech

The nemotron-speech skill is a single routing entry point for all NVIDIA Nemotron Speech (Riva) NIM workflows, covering automatic speech recognition (ASR), text-to-speech (TTS), and neural machine translation (NMT). It solves the problem of navigating NVIDIA's sprawling Riva tooling by acting as a thin routing surface that identifies a user's task type and loads only the relevant detailed reference file, keeping context tight through progressive disclosure. The skill notes that "Nemotron Speech" is a brand-only rename of what NVIDIA still documents as Riva, so all commands, images, gRPC APIs, and doc URLs continue to use the Riva name.

Its capabilities span cloud-hosted inference via build.nvidia.com, self-hosted Docker deployment, client-protocol selection for ASR (gRPC, HTTP, WebSocket), custom NeMo model deployment through the four-phase riva-build / riva-deploy pipeline, ASR pipeline tuning (VAD, diarization, language models), and prerequisite driver/Docker/NGC setup. Reference files cover model selection across Parakeet, Canary, Whisper, Nemotron ASR Streaming, and Magpie TTS, plus deployment-readiness and GPU compatibility checks. The skill is explicitly security-conscious: it instructs treating NVIDIA_API_KEY and NGC_API_KEY as secrets, using --password-stdin for Docker login, storing keys in a credential manager or chmod 600 file, and avoiding world-writable cache directories.

Target users are ML engineers and DevOps practitioners deploying NVIDIA speech models, whether via cloud NIM endpoints or self-hosted GPU infrastructure. Typical use cases include standing up an ASR/TTS/NMT NIM, building a custom fine-tuned ASR pipeline from a .nemo or .riva checkpoint, checking GPU/driver prerequisites, and choosing the right model family for a task.