Nemotron Speech NIM Skill

The skill explicitly prohibits exposing NVIDIA_API_KEY and NGC_API_KEY, recommends non-printing secret checks, and discloses that Docker environment variables may appear in inspect output or process listings. It also routes users to canonical NVIDIA documentation and identifies some permission risks. Points are deducted because the workflow still relies broadly on sudo, Docker, external networks, and remote APIs without a consistently enforced confirmation gate, rollback plan, complete data-flow boundaries, or comprehensive dependency-security guidance; repository license metadata is NOASSERTION while the file claims Apache-2.0.

The documentation provides clear routing, deployment phases, readiness checks, audio validation, and several diagnosable failure messages, while directing release-sensitive values to current NVIDIA documentation. Points are deducted because this review is static and cannot reproduce key paths; cloud functions, container tags, client scripts, and referenced files depend on external state, and command/version compatibility assumptions were not executed or verified. The score therefore stays below the static ceiling of 10.

The audience, ASR/TTS/NMT scenarios, cloud versus self-hosted paths, trigger terms, non-fit ranges, and x86_64/WSL2 limitations are described, and the evaluation material specifies several negative-trigger cases. Points are deducted because generic terms such as Whisper can still cause false activation; Chinese-language support and reachability from mainland-China networks are not substantively demonstrated, while the cloud path depends on build.nvidia.com, NVCF, and GitHub services that may be difficult to reach.

The skill uses a top-level routing surface with progressive disclosure through reference files and includes a version, team metadata, tags, examples, troubleshooting, limitations, and an update path through official documentation. Points are deducted for the absence of a clearly named maintainer, changelog, pinned dependency versions, and complete installation governance; the supplied benchmark also reports documentation-quality issues involving script tables, run_script guidance, and reference organization, while repository license metadata remains unclear.

The material covers model selection, environment checks, deployment, readiness verification, inference, and custom-model conversion, with commands that are often directly adaptable and reduce manual discovery work. Points are deducted because model names, function IDs, container tags, VRAM requirements, and feature support must be re-queried, and no representative outputs were executed or independently verified in this review. Users still bear external-service, credential, GPU, and release-compatibility costs, so the score remains near the static ceiling.

The skill provides many official NVIDIA documentation, API, NGC, and GitHub reference points, plus evaluation data, test tasks, and expected behaviors that offer useful audit trails. Points are deducted because the benchmark and percentages are claims contained in the supplied files, not independently reproducible CI evidence or third-party results; no commands were executed and no cross-source verification was performed. The score therefore stays below the static ceiling of 5.