What problem does it solve? Benchmarking automatic speech recognition models on NVIDIA GPUs requires manual container setup, dataset mounting, inference execution, and metric collection. This Skill automates the full evaluation workflow for the SenseVoiceSmall multilingual ASR model, producing standardized accuracy and throughput reports. ## Core Features & Use Cases - Containerized Evaluation Workflow: Guides Docker container startup with GPU mounting, volume mapping for model weights, datasets, and configs, then runs inference via infer_runner.py. - Multi-Dataset ASR Benchmarking: Supports 28 speech recognition datasets including Aishell-1, LibriSpeech, FLEURS, and KeSpeech across Chinese, English, Japanese, Cantonese, and Korean. - Automated Metric Collection: Extracts CER (character error rate), WER (word error rate), average inference time, and success rate from acc_report.json into a structured result.json. - Use Case: A hardware evaluation engineer needs to measure SenseVoiceSmall recognition accuracy and per-sample latency on an NVIDIA H200 across the KeSpeech dialect datasets; the Skill drives container launch, inference, and metric aggregation end to end. ## Quick Start Ask the agent to test SenseVoiceSmall speech recognition inference performance on NVIDIA GPU using the Aishell-1 dataset and collect the CER metrics.
