New benchmarks (i.e. Nvidia canary) in HuggingFace ASR leaderboard (huggingface.co)

Hugging Face’s open ASR leaderboard (the Spaces project open_asr_leaderboard) has picked up new benchmark entries, including an Nvidia “canary” build, according to recent Spaces activity logs. The logs show the Spaces container fetching metadata from the Hugging Face Docker registry and running the leaderboard app (notably on CPU in the captured runs), which indicates containerized, reproducible benchmark runs are being used to add and validate new entries. The “canary” tag suggests these are pre-release or experimental Nvidia builds being evaluated alongside community models. This matters because adding vendor canaries and other new entries gives the ASR community early, comparable metrics for emerging inference optimizations and model builds. Technical implications include easier side-by-side comparisons of latency, throughput, and accuracy across hardware/software stacks (CPU vs GPU, vendor-specific runtimes), and more transparent reproducibility thanks to HF Docker images and Spaces hosting. Developers and deployers should watch these entries for signs of performance wins (or regressions) from experimental Nvidia changes and expect iterative updates as canary builds mature into stable releases.