import gradio as gr import pandas as pd banner_url = "https://huggingface.co/spaces/elmresearchcenter/open_universal_arabic_asr_leaderboard/resolve/main/banner.png" BANNER = f'
Banner
' INTRODUCTION_TEXT = """ πŸ“–**Open Universal Arabic ASR Leaderboard**πŸ“– benchmarks multi-dialect Arabic ASR models on various multi-dialect datasets. \nApart from the WER%/CER% for each test set, we also report the Average WER%/CER% and rank the models based on the Average WER, from lowest to highest. \nTo reproduce the benchmark numbers and request a model that is not listed, you can launch an issue/PR in our [GitHub repo](https://github.com/Natural-Language-Processing-Elm/open_universal_arabic_asr_leaderboard)😊. \nFor more detailed analysis such as models' robustness, speaker adaption, model efficiency and memory usage, please check our [paper](https://arxiv.org/pdf/2412.13788). """ CITATION_BUTTON_TEXT = """ @article{wang2024open, title={Open Universal Arabic ASR Leaderboard}, author={Wang, Yingzhi and Alhmoud, Anas and Alqurishi, Muhammad}, journal={arXiv preprint arXiv:2412.13788}, year={2024} } """ METRICS_TAB_TEXT = METRICS_TAB_TEXT = """ ## Metrics We report both the Word Error Rate (WER) and Character Error Rate (CER) metrics. ## Reproduction The Open Universal Arabic ASR Leaderboard will be a continuous benchmark project. \nWe open-source the evaluation scripts at our [GitHub repo](https://github.com/Natural-Language-Processing-Elm/open_universal_arabic_asr_leaderboard). \nPlease launch a discussion in our GitHub repo to let us know if you want to learn about the performance of a new model. ## Benchmark datasets | Test Set | Num Dialects | Test (h) | |-------------------------------------------------------------------------------------------------|----------------|-------------| | [SADA](https://www.kaggle.com/datasets/sdaiancai/sada2022) | 10 | 10.7 | | [Common Voice 18.0](https://commonvoice.mozilla.org/en/datasets) | 25 | 12.6 | | [MASC (Clean-Test)](https://ieee-dataport.org/open-access/masc-massive-arabic-speech-corpus) | 7 | 10.5 | | [MASC (Noisy-Test)](https://ieee-dataport.org/open-access/masc-massive-arabic-speech-corpus) | 8 | 14.9 | | [MGB-2](http://www.mgb-challenge.org/MGB-2.html) | Unspecified | 9.6 | ## In-depth Analysis We also provide a comprehensive analysis of the models' robustness, speaker adaptation, inference efficiency and memory consumption. \nPlease check our [paper](https://arxiv.org/pdf/2412.13788) to learn more. """ def styled_message(message): return f"

{message}

" LAST_UPDATED = "NOV 6th 2024" results = { "Model": ["nvidia-conformer-large-arabic (lm)", "nvidia-conformer-large-arabic (greedy)", "openai/whisper-large-v3", "facebook/seamless-m4t-v2-large", "openai/whisper-large-v3-turbo", "openai/whisper-large-v2", "openai/whisper-large", "asafaya/hubert-large-arabic-transcribe/", "openai/whisper-medium", "facebook/mms-1b-all", "openai/whisper-small", "whitefox123/w2v-bert-2.0-arabic-4", "jonatasgrosman/wav2vec2-large-xlsr-53-arabic", "speechbrain/asr-wav2vec2-commonvoice-14-ar"], "Average WER⬇️": [25.71, 27.46, 29.87, 32.55, 33.30, 34.04, 36.65, 39.29, 39.60, 47.86, 48.62, 52.18, 54.63, 60.15], "Average CER": [10.02, 9.94, 13.65, 14.47, 15.68, 16.26, 17.44, 13.61, 19.10, 17.66, 16.79, 25.15, 21.46, 26.64], "SADA WER": [44.52, 47.26, 55.96, 62.52, 60.36, 57.46, 63.24, 67.82, 67.71, 77.48, 78.02, 87.34, 86.82, 88.54], "SADA CER": [23.76, 22.54, 34.62, 37.61, 37.67, 36.59, 40.16, 31.83, 43.83, 37.50, 33.17, 56.75, 44.20, 50.28], "Common Voice WER": [8.80, 10.60, 17.83, 21.70, 25.73, 21.77, 26.04, 8.01, 28.07, 26.52, 24.18, 41.79, 23.00, 29.17], "Common Voice CER": [2.77, 3.05, 5.74, 6.24, 10.89, 7.44, 9.61, 2.37, 10.38, 7.21, 6.79, 15.75, 6.64, 9.85], "MASC(clean-test) WER": [23.74, 24.12, 24.66, 25.04, 25.51, 27.25, 28.89, 32.94, 29.99, 38.82, 35.93, 37.82, 42.75, 49.10], "MASC(clean-test) CER": [5.63, 5.63, 7.24, 7.19, 7.55, 8.28, 9.05, 7.15, 8.98, 10.36, 9.01, 11.92, 11.87, 16.37], "MASC(noisy-test) WER": [34.29, 35.64, 34.63, 33.24, 37.16, 38.55, 40.79, 50.16, 42.91, 57.33, 56.36, 53.28, 64.27, 69.57], "MASC(noisy-test) CER": [11.07, 11.02, 12.89, 11.92, 13.93, 15.49, 16.31, 15.62, 17.49, 19.76, 19.43, 21.93, 24.17, 30.17], "MGB-2 WER": [17.20, 19.69, 16.26, 20.23, 17.75, 25.17, 24.28, 37.51, 29.32, 39.16, 48.64, 40.66, 56.29, 64.37], "MGB-2 CER": [6.87, 7.46, 7.74, 9.37, 8.34, 13.48, 12.10, 11.07, 14.82, 13.48, 15.56, 19.39, 20.44, 26.56], } original_df = pd.DataFrame(results) original_df.sort_values(by="Average WER⬇️", inplace=True) TYPES = ['str', 'number', 'number', 'number', 'number', 'number', 'number', 'number', 'number', 'number', 'number', 'number', 'number'] def request_model(model_text): return styled_message("πŸ€— Please launch a discussion in our GitHub repo, thank you. πŸ€—") with gr.Blocks() as demo: gr.HTML(BANNER, elem_id="banner") gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text") with gr.Tabs(elem_classes="tab-buttons") as tabs: with gr.TabItem("πŸ… Leaderboard", elem_id="od-benchmark-tab-table", id=0): leaderboard_table = gr.Dataframe( value=original_df, datatype=TYPES, elem_id="leaderboard-table", interactive=False, visible=True, ) with gr.TabItem("πŸ“ˆ Metrics", elem_id="od-benchmark-tab-table", id=1): gr.Markdown(METRICS_TAB_TEXT, elem_classes="markdown-text") with gr.TabItem("βœ‰οΈβœ¨ Request a model here!", elem_id="od-benchmark-tab-table", id=2): with gr.Column(): gr.Markdown("# βœ‰οΈβœ¨ Request results for a new model here!", elem_classes="markdown-text") model_name_textbox = gr.Textbox(label="Model name (user_name/model_name)") mdw_submission_result = gr.Markdown() btn_submit = gr.Button(value="πŸš€ Request") btn_submit.click(request_model, [model_name_textbox], mdw_submission_result) gr.Markdown(f"Last updated on **{LAST_UPDATED}**", elem_classes="markdown-text") with gr.Row(): with gr.Accordion("πŸ“™ Citation", open=False): gr.Textbox( value=CITATION_BUTTON_TEXT, lines=7, label="Copy the BibTeX snippet to cite this source", elem_id="citation-button", show_copy_button=True, ) demo.launch(allowed_paths=["banner.png"])