app.py

import os
import gradio as gr
from openai import OpenAI
from jiwer import wer
from rouge_score import rouge_scorer
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity

# Connect to OpenAI API
def get_client():
    api_key = os.getenv("OPENAI_API_KEY")
    if not api_key:
        raise gr.Error("Missing OPENAI_API_KEY. Please set it in the Space Secrets.")
    return OpenAI(api_key=api_key)

# Podcast-style summary prompt
SUMMARY_PROMPT = """
You are a skilled voice script writer. Convert the following lecture transcript into a speech-friendly, podcast-style script suitable for a 3–5 minute audio revision.
- Target audience is already familiar with the video and wants a clear, efficient recap.
- Preserve all key knowledge nodes and insights; do not omit or add content.
- Remove fillers, repetition, and references to slides or visuals.
- Use natural spoken language suitable for listening.
- Maintain a neutral, engaging tone.
- Format as a smooth podcast monologue.
Important Guidelines:
- The summary should be ~20% of the transcript length.
- Do not impersonate or claim to be a real professor or individual.
- Avoid mentioning specific universities, brands, or affiliations unless explicitly present.
- Do not fabricate facts, examples, or names not in the original transcript.
- Ensure all information remains faithful to the transcript.
"""

def run_pipeline(transcript_file):
    if transcript_file is None:
        raise gr.Error("Please upload a .txt transcript file.")

    # Read transcript
    with open(transcript_file.name, "r", encoding="utf-8") as f:
        transcript = f.read()

    client = get_client()

    # Summarization
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": SUMMARY_PROMPT},
            {"role": "user", "content": transcript}
        ]
    )
    script_text = response.choices[0].message.content

    # TTS
    audio_file_path = "summary_audio.mp3"
    tts_response = client.audio.speech.create(
        model="gpt-4o-mini-tts",
        voice="alloy",
        input=script_text
    )
    with open(audio_file_path, "wb") as f:
        f.write(tts_response.read())

    # ASR
    with open(audio_file_path, "rb") as f:
        asr_response = client.audio.transcriptions.create(
            model="whisper-1",
            file=f
        )
    asr_text = asr_response.text.strip()

    # Evaluation
    wer_score = wer(script_text.lower(), asr_text.lower())
    scorer = rouge_scorer.RougeScorer(['rougeL'], use_stemmer=True)
    rouge_l_score = scorer.score(transcript, asr_text)['rougeL'].fmeasure
    vec = TfidfVectorizer().fit_transform([transcript, asr_text])
    cos_sim = cosine_similarity(vec[0:1], vec[1:2])[0][0]

    # Thresholds
    pass_wer = wer_score <= 0.15
    pass_rouge = rouge_l_score >= 0.20
    pass_cosine = cos_sim >= 0.35
    overall_pass = pass_wer and pass_rouge and pass_cosine

    eval_dict = {
        "WER": round(wer_score, 4),
        "WER_pass": pass_wer,
        "ROUGE-L_F1": round(rouge_l_score, 4),
        "ROUGE_pass": pass_rouge,
        "TFIDF_Cosine": round(cos_sim, 4),
        "Cosine_pass": pass_cosine,
        "Overall": "PASS" if overall_pass else "FAIL"
    }

    return script_text, audio_file_path, asr_text, eval_dict


# Build Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("# Transcript → Podcast Summary → TTS → ASR → Evaluation")

    infile = gr.File(label="Upload Transcript (.txt)", file_types=[".txt"])
    run_btn = gr.Button("Run Pipeline")

    summary_out = gr.Textbox(label="Podcast-style Summary", lines=14)
    audio_out = gr.Audio(label="Summary Audio", type="filepath")
    asr_out = gr.Textbox(label="ASR Transcript", lines=10)
    metrics_out = gr.JSON(label="Evaluation Metrics")

    run_btn.click(run_pipeline, inputs=[infile],
                  outputs=[summary_out, audio_out, asr_out, metrics_out])

if __name__ == "__main__":
    demo.launch()