What problem does it solve? Running speech recognition model benchmarks on Ascend NPU hardware requires manual container setup, dataset mounting, inference execution, and metric collection, which is error-prone and time-consuming. This Skill automates the full evaluation workflow for SenseVoice ASR models on Ascend 910B. ## Core Features & Use Cases - Automated Container Orchestration: Launches a preconfigured Docker container with NPU device passthrough, driver mounts, and volume mappings for models, datasets, and configs. - Multi-Dataset ASR Evaluation: Runs SenseVoiceSmall inference across 28 speech recognition datasets including Aishell-1, LibriSpeech, FLEURS, and KeSpeech variants. - Accuracy and Performance Metrics: Collects CER (character error rate) for Chinese datasets, WER (word error rate) for English datasets, and average inference time per sample. - Use Case: An engineer needs to validate SenseVoiceSmall recognition accuracy and throughput on Ascend 910B before production deployment; this Skill runs the full benchmark and produces a structured acc_report.json. ## Quick Start Ask the agent to test SenseVoiceSmall speech recognition inference performance on Ascend NPU using the Aishell-1 dataset and report CER and average inference time.
