app.py

# imports 
import gradio as gr
import os
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 = {
    "hausa": "asr-africa/wav2vec2-xls-r-1b-naijavoices-hausa-500hr-v0",
    "igbo": "asr-africa/wav2vec2-xls-r-1b-naijavoices-igbo-500hr-v0",
    "yoruba": "asr-africa/wav2vec2-xls-r-1b-naijavoices-yoruba-500hr-v0",
    "zulu": "asr-africa/W2V2-Bert_nchlt_speech_corpus_Fleurs_ZULU_63hr_v1",
    "xhosa": "asr-africa/wav2vec2_xls_r_300m_nchlt_speech_corpus_Fleurs_XHOSA_63hr_v1",
    "afrikaans": "asr-africa/mms-1B_all_nchlt_speech_corpus_Fleurs_CV_AFRIKAANS_57hr_v1",
    "bemba": "asr-africa/whisper_BIG-C_BEMBA_189hr_v1",
    "shona": "asr-africa/W2V2_Bert_Afrivoice_FLEURS_Shona_100hr_v1",
    "luganda": "asr-africa/whisper-small-CV-Fleurs-lg-313hrs-v1",
    "swahili": "asr-africa/wav2vec2-xls-r-300m-CV_Fleurs_AMMI_ALFFA-sw-400hrs-v1",
    "lingala": "asr-africa/wav2vec2-xls-r-300m-Fleurs_AMMI_AFRIVOICE_LRSC-ln-109hrs-v2",
    "amharic": "asr-africa/facebook-mms-1b-all-common_voice_fleurs-amh-200hrs-v1",
    "kinyarwanda": "asr-africa/facebook-mms-1b-all-common_voice_fleurs-rw-100hrs-v1",
    "oromo": "asr-africa/mms-1b-all-Sagalee-orm-85hrs-4",
    "akan": "asr-africa/wav2vec2-xls-r-akan-100-hours",
    "ewe": "asr-africa/wav2vec2-xls-r-ewe-100-hours",
    "wolof": "asr-africa/w2v2-bert-Wolof-20-hours-Google-Fleurs-ALF-dataset",
    "bambara": "asr-africa/mms-bambara-50-hours-mixed-bambara-dataset",
}

# 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"]
    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")
        }

        response = requests.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)
    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")
    transcribed_text = gr.Textbox(label="Transcription", 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, 10, 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(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()