from pathlib import Path import gradio as gr from aip_trainer import PROJECT_ROOT_FOLDER, app_logger, sample_rate_start from aip_trainer.lambdas import js, lambdaGetSample, lambdaSpeechToScore, lambdaTTS css = """ .speech-output-label p {color: grey; margin-bottom: white;} .background-white {background-color: white !important; } .speech-output-group {padding: 12px;} .speech-output-container {min-height: 60px;} .speech-output-html {text-align: left; } """ def clear(): return None def clear2(): return None, None with gr.Blocks(css=css, head=js.head_driver_tour) as gradio_app: local_storage = gr.BrowserState([0.0, 0.0]) app_logger.info("start gradio app building...") project_root_folder = Path(PROJECT_ROOT_FOLDER) with open(project_root_folder / "aip_trainer" / "lambdas" / "app_description.md", "r", encoding="utf-8") as app_description_src: md_app_description = app_description_src.read() gr.Markdown(md_app_description.format(sample_rate_start=sample_rate_start)) with gr.Row(): with gr.Column(scale=4, min_width=300): with gr.Row(): with gr.Column(scale=2, min_width=80): radio_language = gr.Radio(["de", "en"], label="Language", value="en", elem_id="radio-language-id-element") with gr.Column(scale=5, min_width=160): radio_difficulty = gr.Radio( label="Difficulty", value=0, choices=[ ("random", 0), ("easy", 1), ("medium", 2), ("hard", 3), ], elem_id="radio-difficulty-id-element", ) with gr.Column(scale=1, min_width=100): btn_random_phrase = gr.Button(value="Choose a random phrase", elem_id="btn-random-phrase-id-element") with gr.Row(): with gr.Column(scale=7, min_width=300): text_student_transcription = gr.Textbox( lines=3, label="Phrase to read for speech recognition", value="Hi there, how are you?", elem_id="text-student-transcription-id-element", ) with gr.Row(): audio_tts = gr.Audio(label="Audio TTS", elem_id="audio-tts-id-element") with gr.Row(): btn_run_tts = gr.Button(value="TTS in browser", elem_id="btn-run-tts-id-element") btn_run_tts_backend = gr.Button(value="TTS backend", elem_id="btn-run-tts-backend-id-element") btn_clear_tts = gr.Button(value="Clear TTS backend", elem_id="btn-clear-tts-backend-id-element") btn_clear_tts.click(clear, inputs=[], outputs=[audio_tts]) with gr.Row(): audio_student_recording_stt = gr.Audio( label="Speech-toText audio output", sources=["microphone", "upload"], type="filepath", show_download_button=True, elem_id="audio-student-recording-stt-id-element", ) with gr.Column(scale=4, min_width=320): text_transcribed_hidden = gr.Textbox( placeholder=None, label="Transcribed text", visible=False ) text_letter_correctness = gr.Textbox( placeholder=None, label="Letters correctness", visible=False, ) text_recording_ipa = gr.Textbox( placeholder=None, label="Student phonetic transcription", elem_id="text-student-recording-ipa-id-element" ) text_ideal_ipa = gr.Textbox( placeholder=None, label="Ideal phonetic transcription", elem_id="text-ideal-ipa-id-element" ) text_raw_json_output_hidden = gr.Textbox(placeholder=None, label="text_raw_json_output_hidden", visible=False) with gr.Group(elem_classes="speech-output-group background-white"): gr.Markdown("Speech accuracy output", elem_classes="speech-output-label background-white") with gr.Group(elem_classes="speech-output-container background-white"): html_output = gr.HTML( label="Speech accuracy output", elem_id="speech-output", show_label=False, visible=True, render=True, value=" - ", elem_classes="speech-output-html background-white", ) with gr.Row(): gr.Markdown("### Speech accuracy score (%)", elem_classes="speech-accuracy-score-container row1", elem_id="speech-accuracy-score-container-id-element") with gr.Row(): with gr.Column(min_width=100, elem_classes="speech-accuracy-score-container row2 col1"): number_pronunciation_accuracy = gr.Number(label="Current score", elem_id="number-pronunciation-accuracy-id-element") with gr.Column(min_width=100, elem_classes="speech-accuracy-score-container row2 col2"): number_score_de = gr.Number(label="Global score DE", value=0, interactive=False, elem_id="number-score-de-id-element") with gr.Column(min_width=100, elem_classes="speech-accuracy-score-container row2 col3"): number_score_en = gr.Number(label="Global score EN", value=0, interactive=False, elem_id="number-score-en-id-element") with gr.Row(): btn = gr.Button(value="Recognize speech accuracy", elem_id="btn-recognize-speech-accuracy-id-element") with gr.Accordion("Click here to expand the table examples", open=True, elem_id="accordion-examples-id-element"): examples_text = gr.Examples( examples=[ ["Hallo, wie geht es dir?", "de", 1], ["Hi there, how are you?", "en", 1], ["Die König-Ludwig-Eiche ist ein Naturdenkmal im Staatsbad Brückenau.", "de", 2,], ["Rome is home to some of the most beautiful monuments in the world.", "en", 2], ["Die König-Ludwig-Eiche ist ein Naturdenkmal im Staatsbad Brückenau, einem Ortsteil des drei Kilometer nordöstlich gelegenen Bad Brückenau im Landkreis Bad Kissingen in Bayern.", "de", 3], ["Some machine learning models are designed to understand and generate human-like text based on the input they receive.", "en", 3], ], inputs=[text_student_transcription, radio_language, radio_difficulty], elem_id="examples-text-id-element", ) def get_updated_score_by_language(text: str, audio_rec: str | Path, lang: str, score_de: float, score_en: float): _transcribed_text, _letter_correctness, _pronunciation_accuracy, _recording_ipa, _ideal_ipa, _res = lambdaSpeechToScore.get_speech_to_score_tuple(text, audio_rec, lang) output = { text_transcribed_hidden: _transcribed_text, text_letter_correctness: _letter_correctness, number_pronunciation_accuracy: _pronunciation_accuracy, text_recording_ipa: _recording_ipa, text_ideal_ipa: _ideal_ipa, text_raw_json_output_hidden: _res, } match lang: case "de": return { number_score_de: float(score_de) + float(_pronunciation_accuracy), number_score_en: float(score_en), **output } case "en": return { number_score_en: float(score_en) + float(_pronunciation_accuracy), number_score_de: float(score_de), **output } case _: raise NotImplementedError(f"Language {lang} not supported") btn.click( get_updated_score_by_language, inputs=[text_student_transcription, audio_student_recording_stt, radio_language, number_score_de, number_score_en], outputs=[ text_transcribed_hidden, text_letter_correctness, number_pronunciation_accuracy, text_recording_ipa, text_ideal_ipa, text_raw_json_output_hidden, number_score_de, number_score_en ], ) btn_run_tts.click(fn=None, inputs=[text_student_transcription, radio_language], outputs=audio_tts, js=js.js_play_audio) btn_run_tts_backend.click( fn=lambdaTTS.get_tts, inputs=[text_student_transcription, radio_language], outputs=audio_tts, ) btn_random_phrase.click( lambdaGetSample.get_random_selection, inputs=[radio_language, radio_difficulty], outputs=[text_student_transcription], ) btn_random_phrase.click( clear2, inputs=[], outputs=[audio_student_recording_stt, audio_tts] ) html_output.change( None, inputs=[text_transcribed_hidden, text_letter_correctness], outputs=[html_output], js=js.js_update_ipa_output, ) @gradio_app.load(inputs=[local_storage], outputs=[number_score_de, number_score_en]) def load_from_local_storage(saved_values): print("loading from local storage", saved_values) return saved_values[0], saved_values[1] @gr.on([number_score_de.change, number_score_en.change], inputs=[number_score_de, number_score_en], outputs=[local_storage]) def save_to_local_storage(score_de, score_en): return [score_de, score_en] if __name__ == "__main__": gradio_app.launch()