# imports import gradio as gr import os import cloudscraper import requests from transformers import pipeline import torch HF_TOKEN = os.getenv("HF_TOKEN") # Set your FastAPI backend endpoint BACKEND_URL = "https://asr-evaluation-backend.emergentai.ug/submit-feedback" model_map = { "afrikaans": "asr-africa/mms-1B_all_nchlt_speech_corpus_Fleurs_CV_AFRIKAANS_57hr_v1", "akan": "asr-africa/wav2vec2-xls-r-akan-100-hours", "amharic": "asr-africa/facebook-mms-1b-all-common_voice_fleurs-amh-200hrs-v1", "bambara": "asr-africa/mms-bambara-50-hours-mixed-bambara-dataset", "bemba": "asr-africa/whisper_BIG-C_BEMBA_189hr_v1", "ewe": "asr-africa/wav2vec2-xls-r-ewe-100-hours", "hausa": "asr-africa/wav2vec2-xls-r-1b-naijavoices-hausa-500hr-v0", "igbo": "asr-africa/wav2vec2-xls-r-1b-naijavoices-igbo-500hr-v0", "kinyarwanda": "asr-africa/facebook-mms-1b-all-common_voice_fleurs-rw-100hrs-v1", "lingala": "asr-africa/wav2vec2-xls-r-300m-Fleurs_AMMI_AFRIVOICE_LRSC-ln-109hrs-v2", "luganda": "asr-africa/whisper-small-CV-Fleurs-lg-313hrs-v1", "oromo": "asr-africa/mms-1b-all-Sagalee-orm-85hrs-4", "shona": "asr-africa/W2V2_Bert_Afrivoice_FLEURS_Shona_100hr_v1", "swahili": "asr-africa/wav2vec2-xls-r-300m-CV_Fleurs_AMMI_ALFFA-sw-400hrs-v1-nolm", "wolof": "asr-africa/w2v2-bert-Wolof-20-hours-Google-Fleurs-ALF-dataset", "xhosa": "asr-africa/wav2vec2_xls_r_300m_nchlt_speech_corpus_Fleurs_XHOSA_63hr_v1", "yoruba": "asr-africa/wav2vec2-xls-r-1b-naijavoices-yoruba-500hr-v0", "zulu": "asr-africa/W2V2-Bert_nchlt_speech_corpus_Fleurs_ZULU_63hr_v1", } # Create storage directory os.makedirs("responses", exist_ok=True) # Transcription function inference_device = 0 if torch.cuda.is_available() else -1 def transcribe(audio, language): asr = pipeline("automatic-speech-recognition", model=model_map[language], device=inference_device, token=HF_TOKEN) text = asr(audio)["text"] # handling whisper models return text, audio # Save feedback by sending it to FastAPI backend def save_feedback(audio_file, transcription, user_id, lang, env, device, domain, accuracy, transcript_edit, orthography, orthography_issues, meaning, meaning_loss, errors, error_examples, performance): try: with open(audio_file, "rb") as f: audio_content = f.read() metadata = { "transcription": transcription, "user_id": user_id, "transcript_edit": transcript_edit, "evaluated_language": lang, "environment": env, "device": device, "domain": domain, "accuracy": accuracy, "orthography": orthography, "orthography_issues": orthography_issues, "meaning": meaning, "meaning_loss": meaning_loss, "errors": ",".join(errors) if errors else "", "error_examples": error_examples, "performance": performance } files = { "audio_file": ("audio.wav", audio_content, "audio/wav") } scraper = cloudscraper.create_scraper() response = scraper.post(BACKEND_URL, data=metadata, files=files, timeout=20) if response.status_code == 201: return "✅ Feedback submitted successfully. Thank you!" else: return f"⚠️ Submission failed: {response.status_code} — {response.text}" except Exception as e: return f"❌ Could not connect to the backend: {str(e)}" # Gradio UI with gr.Blocks() as demo: gr.Markdown("## African ASR Evaluation Platform") gr.Markdown("**Select Language**") lang = gr.Dropdown(list(model_map.keys()), label="", value=None) gr.Markdown("**Upload or Record Audio**") audio_input = gr.Audio(sources=["upload", "microphone"], type="filepath", label="Upload or record audio") # transcribed_text = gr.Textbox(label="Transcription", interactive=False) submit_btn = gr.Button("Transcribe") gr.Markdown("**Transcription**") transcribed_text = gr.Textbox(label="", interactive=False) submit_btn.click(fn=transcribe, inputs=[audio_input, lang], outputs=[transcribed_text, audio_input]) gr.Markdown("---\n## Feedback Form") user_id = gr.Textbox(label="Please enter user ID.*") env = gr.Dropdown(["Studio/Professional Recording", "Quiet Room (minimal noise)", "Noisy Background (e.g., street, cafe, market)"], label="What was the type of recording environment for the speech you evaluated? *",value=None) device = gr.Dropdown(["Mobile Phone/Tablet", "Laptop/Computer Microphone", "Dedicated Microphone (e.g., headset, studio mic)"], label="What type of recording device was used? *",value=None) domain = gr.Textbox(label="Was the speech related to a specific topic? If yes, please specify the topic (e.g., news, education, medical, law, religious, sports, science).") accuracy = gr.Slider(1, 5, step=1, label="Overall, how accurate was the model's transcription for the audio you reviewed? *") transcript_edit = gr.Textbox(label="If the transcription provided by the model was incorrect, please enter your corrected version.") orthography = gr.Radio(["Yes, mostly correct", "No, major issues", "Partially (some correct, some incorrect)", "Not Applicable"], label="Did the transcription correctly use the standard orthography (including accents, diacritics, special characters) for the language?",value=None) orthography_issues = gr.Textbox(label="If you selected \"No\" or \"Partially\", please describe any significant orthography issues you noticed.") meaning = gr.Slider(1, 5, step=1, label="Did the model's transcription preserve the original meaning of the speech? *") meaning_loss = gr.Textbox(label="If the meaning was not fully preserved (i.e., you rated 1-4 above), please briefly explain how it was changed or lost.") errors = gr.CheckboxGroup([ "Substitutions (wrong words used)", "Omissions (words missing)", "Insertions (extra words added)", "Pronunciation-related errors (phonetically plausible but wrong word/spelling)", "Diacritic/Tone/Special Character errors", "Code-switching errors (mixing languages incorrectly)", "Named Entity errors (names of people/places wrong)", "Punctuation errors", "No significant errors observed" ] , label="Which types of errors were most prominent or impactful in the transcriptions? *", value=[]) error_examples = gr.Textbox(label="(Optional) Can you provide 1-2 examples of significant errors and how you would correct them?") performance = gr.Textbox(label="Please describe the model's performance in your own words. What did it do well? What did it struggle with? *") save_btn = gr.Button("Submit Feedback") output_msg = gr.Textbox(label="Submission status",interactive=False) save_btn.click( fn=save_feedback, inputs=[ audio_input, transcribed_text, user_id, lang, env, device, domain, accuracy, transcript_edit, orthography, orthography_issues, meaning, meaning_loss, errors, error_examples, performance ], outputs=[output_msg] ) # Launch the interface demo.launch()