open_asr_leaderboard

Aggregates evaluation metrics (WER, CER, latency) across multiple open-source speech recognition models tested on standardized datasets, then ranks and visualizes results in a sortable leaderboard interface. Uses Hugging Face Model Hub integration to fetch model metadata and evaluation results, with real-time updates as new model submissions are processed through an automated evaluation pipeline.

Executes standardized speech recognition inference on submitted models using a fixed set of test datasets and metrics (WER, CER, latency), then stores results in a structured format for leaderboard ranking. The pipeline likely uses Hugging Face Transformers library to load models, librosa or similar for audio processing, and jiwer or similar for WER computation, with results persisted to a database or JSON store that feeds the leaderboard UI.

Provides a Gradio-based web interface with sortable columns, search functionality, and optional filtering controls to explore the ranked ASR models. Users can click column headers to sort by WER, latency, or other metrics, and may filter by language, model size, or other metadata attributes. The interface is built with Gradio components (Table, Dropdown, Textbox) that bind to backend data structures, enabling real-time sorting without page reloads.

Displays structured metadata for each ranked model (model name, author, language support, model size, architecture type) and provides direct hyperlinks to the model's Hugging Face repository, paper, or demo. Metadata is fetched from model cards stored in the Hub and enriched with evaluation results, creating a unified view that connects leaderboard rankings to source code, weights, and documentation.

Renders performance metrics (WER, latency, model size) in visual formats such as scatter plots, bar charts, or heatmaps to help users understand accuracy-speed-size tradeoffs across models. Likely uses Plotly or similar charting library integrated with Gradio to generate interactive visualizations that update when users filter or sort the leaderboard, enabling quick visual identification of Pareto-optimal models.