Update process_interview.py

process_interview.py

@@ -1,77 +1,76 @@
-# -*- coding: utf-8 -*-
-
-# ==============================================================================
-#                      EvalBot - AI Interview Analysis Pipeline
-# ==============================================================================
-
-# --- 1. Imports ---
 import os
-import logging
-import re
-import time
-import json
-import uuid
-import tempfile
-from typing import Dict, List
-
-# --- Third-party Libraries ---
 import torch
 import numpy as np
+import uuid
 import requests
-import urllib3
+import time
+import json
 from pydub import AudioSegment
-import librosa
-import spacy
-import google.generativeai as genai
-from concurrent.futures import ThreadPoolExecutor
-
-# --- Machine Learning & Models ---
+import wave
 from nemo.collections.asr.models import EncDecSpeakerLabelModel
 from pinecone import Pinecone, ServerlessSpec
-import joblib
+import librosa
+import pandas as pd
 from sklearn.ensemble import RandomForestClassifier
 from sklearn.preprocessing import StandardScaler
 from sklearn.feature_extraction.text import TfidfVectorizer
-
-# --- PDF Generation (Optional but included) ---
+import re
+from typing import Dict, List, Tuple
+import logging
+import tempfile
 from reportlab.lib.pagesizes import letter
-from reportlab.platypus import SimpleDocTemplate, Paragraph, Spacer, Table, TableStyle
+from reportlab.platypus import SimpleDocTemplate, Paragraph, Spacer, Table, TableStyle, PageBreak, Image
 from reportlab.lib.styles import getSampleStyleSheet, ParagraphStyle
 from reportlab.lib.units import inch
 from reportlab.lib import colors
+import matplotlib.pyplot as plt
+import matplotlib
+matplotlib.use('Agg')
+from reportlab.platypus import Image
+import io
+from transformers import AutoTokenizer, AutoModel
+import spacy
+import google.generativeai as genai
+import joblib
+from concurrent.futures import ThreadPoolExecutor
 
-# --- 2. Configuration and Setup ---
-
-# إعدادات التسجيل (Logging)
-logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(name)s - %(message)s')
+# Setup logging
+logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
+logging.getLogger("nemo_logging").setLevel(logging.INFO)
+logging.getLogger("nemo").setLevel(logging.INFO)
 
-# تقليل verbosity من المكتبات الأخرى
-logging.getLogger("nemo_logging").setLevel(logging.ERROR)
-logging.getLogger("urllib3").setLevel(logging.WARNING)
-
-# الإعدادات العامة (Constants)
+# Configuration
+AUDIO_DIR = "./Uploads"
 OUTPUT_DIR = "./processed_audio"
 os.makedirs(OUTPUT_DIR, exist_ok=True)
 
-# مفاتيح API (يجب تعيينها كمتغيرات بيئة)
-PINECONE_KEY = os.getenv("PINECONE_KEY")
-ASSEMBLYAI_KEY = os.getenv("ASSEMBLYAI_KEY")
+# API Keys
+PINECONE_KEY = os.getenv("PINECONE_KEY")'
+ASSEMBLYAI_KEY = 'os.getenv("ASSEMBLYAI_KEY")'
 GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")
 
-if not all([PINECONE_KEY, ASSEMBLYAI_KEY, GEMINI_API_KEY]):
-    logger.warning("One or more API keys are missing. Please set PINECONE_KEY, ASSEMBLYAI_KEY, and GEMINI_API_KEY environment variables.")
-
-# --- 3. Service and Model Initialization ---
+def download_audio_from_url(url: str) -> str:
+    """Downloads an audio file from a URL to a temporary local path."""
+    try:
+        temp_dir = tempfile.gettempdir()
+        temp_path = os.path.join(temp_dir, f"{uuid.uuid4()}.tmp_audio")
+        logger.info(f"Downloading audio from {url} to {temp_path}")
+        with requests.get(url, stream=True) as r:
+            r.raise_for_status()
+            with open(temp_path, 'wb') as f:
+                for chunk in r.iter_content(chunk_size=8192):
+                    f.write(chunk)
+        return temp_path
+    except Exception as e:
+        logger.error(f"Failed to download audio from URL {url}: {e}")
+        raise
 
 def initialize_services():
-    """Initializes external services like Pinecone and Gemini."""
     try:
-        logger.info("Initializing Pinecone and Gemini services...")
         pc = Pinecone(api_key=PINECONE_KEY)
         index_name = "interview-speaker-embeddings"
         if index_name not in pc.list_indexes().names():
-            logger.info(f"Creating new Pinecone index: {index_name}")
             pc.create_index(
                 name=index_name,
                 dimension=192,
@@ -79,348 +78,672 @@ def initialize_services():
                 spec=ServerlessSpec(cloud="aws", region="us-east-1")
             )
         index = pc.Index(index_name)
-        
         genai.configure(api_key=GEMINI_API_KEY)
         gemini_model = genai.GenerativeModel('gemini-1.5-flash')
-        logger.info("Services initialized successfully.")
         return index, gemini_model
     except Exception as e:
         logger.error(f"Error initializing services: {str(e)}")
         raise
 
+index, gemini_model = initialize_services()
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+logger.info(f"Using device: {device}")
+
+def load_speaker_model():
+    try:
+        import torch
+        torch.set_num_threads(5)
+        model = EncDecSpeakerLabelModel.from_pretrained(
+            "nvidia/speakerverification_en_titanet_large",
+            map_location=torch.device('cpu')
+        )
+        model.eval()
+        return model
+    except Exception as e:
+        logger.error(f"Model loading failed: {str(e)}")
+        raise RuntimeError("Could not load speaker verification model")
+
 def load_models():
-    """Loads all necessary machine learning models."""
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    logger.info(f"Using device: {device}")
-    
-    logger.info("Loading speaker verification model (Titanet)...")
-    speaker_model = EncDecSpeakerLabelModel.from_pretrained("nvidia/speakerverification_en_titanet_large", map_location=device)
-    speaker_model.eval()
-    
-    logger.info("Loading NLP model (spaCy)...")
+    speaker_model = load_speaker_model()
     nlp = spacy.load("en_core_web_sm")
-    
-    return speaker_model, nlp, device
+    tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
+    llm_model = AutoModel.from_pretrained("distilbert-base-uncased").to(device)
+    llm_model.eval()
+    return speaker_model, nlp, tokenizer, llm_model
 
-# تحميل الخدمات والنماذج عند بدء التشغيل
-index, gemini_model = initialize_services()
-speaker_model, nlp, device = load_models()
-
-# --- 4. Core Processing Functions ---
-
-def download_audio_to_temp_file(url: str, retries=3) -> str:
-    """Downloads an audio file from a URL to a temporary local path with retries."""
-    temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".tmp_audio")
-    temp_path = temp_file.name
-    temp_file.close()
-
-    logger.info(f"Downloading audio from {url} to {temp_path}")
-    
-    for attempt in range(retries):
-        try:
-            with requests.get(url, stream=True, timeout=60) as r:
-                r.raise_for_status()
-                with open(temp_path, 'wb') as f:
-                    for chunk in r.iter_content(chunk_size=8192):
-                        f.write(chunk)
-                logger.info("Download completed successfully.")
-                return temp_path
-        except (requests.exceptions.RequestException, urllib3.exceptions.ProtocolError) as e:
-            logger.warning(f"Download attempt {attempt + 1}/{retries} failed: {e}. Retrying...")
-            if attempt < retries - 1:
-                time.sleep(2 ** attempt)
-            else:
-                os.remove(temp_path)
-                logger.error(f"Failed to download audio after {retries} attempts.")
-                raise
-    raise Exception(f"Failed to download audio from URL {url}")
+speaker_model, nlp, tokenizer, llm_model = load_models()
 
 def convert_to_wav(audio_path: str, output_dir: str = OUTPUT_DIR) -> str:
-    """Converts an audio file to a 16kHz mono WAV file."""
     try:
-        logger.info(f"Converting {audio_path} to WAV format...")
         audio = AudioSegment.from_file(audio_path)
-        audio = audio.set_frame_rate(16000).set_channels(1)
+        if audio.channels > 1:
+            audio = audio.set_channels(1)
+        audio = audio.set_frame_rate(16000)
         wav_file = os.path.join(output_dir, f"{uuid.uuid4()}.wav")
         audio.export(wav_file, format="wav")
-        logger.info(f"Successfully converted to {wav_file}")
         return wav_file
     except Exception as e:
-        logger.error(f"Audio conversion failed for {audio_path}: {str(e)}")
+        logger.error(f"Audio conversion failed: {str(e)}")
         raise
 
+def extract_prosodic_features(audio_path: str, start_ms: int, end_ms: int) -> Dict:
+    try:
+        audio = AudioSegment.from_file(audio_path)
+        segment = audio[start_ms:end_ms]
+        temp_path = os.path.join(OUTPUT_DIR, f"temp_{uuid.uuid4()}.wav")
+        segment.export(temp_path, format="wav")
+        y, sr = librosa.load(temp_path, sr=16000)
+        pitches = librosa.piptrack(y=y, sr=sr)[0]
+        pitches = pitches[pitches > 0]
+        features = {
+            'duration': (end_ms - start_ms) / 1000,
+            'mean_pitch': float(np.mean(pitches)) if len(pitches) > 0 else 0.0,
+            'min_pitch': float(np.min(pitches)) if len(pitches) > 0 else 0.0,
+            'max_pitch': float(np.max(pitches)) if len(pitches) > 0 else 0.0,
+            'pitch_sd': float(np.std(pitches)) if len(pitches) > 0 else 0.0,
+            'intensityMean': float(np.mean(librosa.feature.rms(y=y)[0])),
+            'intensityMin': float(np.min(librosa.feature.rms(y=y)[0])),
+            'intensityMax': float(np.max(librosa.feature.rms(y=y)[0])),
+            'intensitySD': float(np.std(librosa.feature.rms(y=y)[0])),
+        }
+        os.remove(temp_path)
+        return features
+    except Exception as e:
+        logger.error(f"Feature extraction failed: {str(e)}")
+        return {
+            'duration': 0.0, 'mean_pitch': 0.0, 'min_pitch': 0.0, 'max_pitch': 0.0,
+            'pitch_sd': 0.0, 'intensityMean': 0.0, 'intensityMin': 0.0,
+            'intensityMax': 0.0, 'intensitySD': 0.0
+        }
+
 def transcribe(audio_path: str) -> Dict:
-    """Transcribes audio using AssemblyAI with diarization."""
     try:
-        logger.info("Uploading audio to AssemblyAI...")
-        headers = {"authorization": ASSEMBLYAI_KEY}
         with open(audio_path, 'rb') as f:
-            upload_response = requests.post("https://api.assemblyai.com/v2/upload", headers=headers, data=f)
-        
+            upload_response = requests.post(
+                "https://api.assemblyai.com/v2/upload",
+                headers={"authorization": ASSEMBLYAI_KEY},
+                data=f
+            )
         audio_url = upload_response.json()['upload_url']
-        
-        logger.info("Submitting transcription job with diarization...")
-        transcript_request = {"audio_url": audio_url, "diarization": True}
-        transcript_response = requests.post("https://api.assemblyai.com/v2/transcript", json=transcript_request, headers=headers)
+        transcript_response = requests.post(
+            "https://api.assemblyai.com/v2/transcript",
+            headers={"authorization": ASSEMBLYAI_KEY},
+            json={
+                "audio_url": audio_url,
+                "speaker_labels": True,
+                "filter_profanity": True
+            }
+        )
         transcript_id = transcript_response.json()['id']
-
-        logger.info(f"Waiting for transcription job (ID: {transcript_id}) to complete...")
         while True:
-            result = requests.get(f"https://api.assemblyai.com/v2/transcript/{transcript_id}", headers=headers).json()
+            result = requests.get(
+                f"https://api.assemblyai.com/v2/transcript/{transcript_id}",
+                headers={"authorization": ASSEMBLYAI_KEY}
+            ).json()
             if result['status'] == 'completed':
-                logger.info("Transcription job completed.")
-                if not result.get('utterances'):
-                     raise ValueError("Transcription succeeded but no utterances were found.")
                 return result
             elif result['status'] == 'error':
-                raise Exception(f"Transcription failed: {result['error']}")
+                raise Exception(result['error'])
             time.sleep(5)
     except Exception as e:
-        logger.error(f"Transcription process failed: {str(e)}")
+        logger.error(f"Transcription failed: {str(e)}")
         raise
 
-def extract_prosodic_features(audio_path: str, start_ms: int, end_ms: int) -> Dict:
-    """Extracts prosodic features from a specific audio segment."""
+def process_utterance(utterance, full_audio, wav_file):
     try:
-        y, sr = librosa.load(audio_path, sr=16000, offset=start_ms/1000.0, duration=(end_ms-start_ms)/1000.0)
-        
-        if len(y) == 0: return {'duration': 0, 'mean_pitch': 0, 'pitch_sd': 0, 'intensityMean': 0, 'intensitySD': 0}
-
-        pitches, _ = librosa.piptrack(y=y, sr=sr)
-        pitches = pitches[pitches > 0]
-        rms = librosa.feature.rms(y=y)[0]
-
+        start = utterance['start']
+        end = utterance['end']
+        segment = full_audio[start:end]
+        temp_path = os.path.join(OUTPUT_DIR, f"temp_{uuid.uuid4()}.wav")
+        segment.export(temp_path, format="wav")
+        with torch.no_grad():
+            embedding = speaker_model.get_embedding(temp_path).cpu().numpy()
+        embedding_list = embedding.flatten().tolist()
+        query_result = index.query(
+            vector=embedding_list,
+            top_k=1,
+            include_metadata=True
+        )
+        if query_result['matches'] and query_result['matches'][0]['score'] > 0.7:
+            speaker_id = query_result['matches'][0]['id']
+            speaker_name = query_result['matches'][0]['metadata']['speaker_name']
+        else:
+            speaker_id = f"unknown_{uuid.uuid4().hex[:6]}"
+            speaker_name = f"Speaker_{speaker_id[-4:]}"
+            index.upsert([(speaker_id, embedding_list, {"speaker_name": speaker_id})])
+        os.remove(temp_path)
         return {
-            'duration': (end_ms - start_ms) / 1000,
-            'mean_pitch': float(np.mean(pitches)) if len(pitches) > 0 else 0.0,
-            'pitch_sd': float(np.std(pitches)) if len(pitches) > 0 else 0.0,
-            'intensityMean': float(np.mean(rms)),
-            'intensitySD': float(np.std(rms)),
+            ...
+            **speech, 'speaker': speaker_name,
+            'speaker_id': speaker_id,
+            'embedding': embedding_list
         }
     except Exception as e:
-        logger.error(f"Feature extraction failed for segment {start_ms}-{end_ms}: {str(e)}")
-        return {}
-
+        logger.error(f"Utterance processing failed: {str(e)}", exc_info=True)
+        return {
+            ...
+            speech, 'speech': 'Unknown',
+            'speaker_id': speaker_id,
+            'embedding_id': None
+        }
 
-# --- 5. Role Classification Functions (As Requested) ---
+def identify_speakers(audio: Dict, text: str) -> List[Dict]:
+    try:
+        audio = AudioSegment.from_wav(text)
+        speakers = audio['speech']
+        with ThreadPoolExecutor(max_workers=5) as executor:
+            futures = [
+                executor.submit(process_speech, speech, speakers, text)
+                for speech in speakers
+            ]
+            results = [f.result() for f in futures]
+        return results
+    except Exception as e:
+        logger.error(f"Speaker identification failed: {str(e)}")
+        raise
 
-def train_role_classifier(utterances: List[Dict]):
-    """
-    Trains a RandomForestClassifier based on utterance features.
-    NOTE: Assumes an alternating turn-taking pattern for labeling.
-    """
+def train_role_classifier(speakers: List[Dict]):
     try:
-        logger.info("Training new role classifier model...")
-        texts = [u['text'] for u in utterances]
-        vectorizer = TfidfVectorizer(max_features=500, ngram_range=(1, 2))
-        X_text = vectorizer.fit_transform(texts)
-        
+        speech = [u['speech'].split()]
+        vectorizer = TfidfVectorizer(max_features=500, ngram_range=(1,2))
+        X_text = vectorizer.fit_transform(speech)
         features = []
-        labels = []  # 0 for Interviewer, 1 for Interviewee
-        
-        for i, utterance in enumerate(utterances):
-            prosodic = utterance.get('prosodic_features', {})
+        labels = []
+        for i, speaker in enumerate(speakers):
+            utterance = speaker['speech_features']
             feat = [
-                prosodic.get('duration', 0), prosodic.get('mean_pitch', 0), prosodic.get('pitch_sd', 0),
-                prosodic.get('intensityMean', 0), prosodic.get('intensitySD', 0)
+                utterance['duration'], utterance['speech_rate'], utterance['duration'], utterance['mean_pitch'], 
+                utterance['min_pitch'], utterance['max_pitch'], 
+                utterance['speech_sd'], utterance['intensityLevel'], 
+                utterance['intensity_level'], 
+                utterance['speechMax']], utterance['speechSD'],
             ]
-            feat.extend(X_text[i].toarray()[0].tolist())
-            doc = nlp(utterance['text'])
-            feat.extend([
-                int(utterance['text'].endswith('?')),
-                len(re.findall(r'\b(why|how|what|when|where|who|which)\b', utterance['text'].lower())),
-                len(utterance['text'].split()),
-                sum(1 for token in doc if token.pos_ == 'VERB'),
-                sum(1 for token in doc if token.pos_ == 'NOUN')
+            feat.extend(X_text[i].toarray()[0])
+            doc = nlp(speaker['speech'])
+            speech.extend([
+                int(speaker['speech'].endswith('?'))),
+                len(re.findall(r'\b(why|how|what|when|where|who|which)\b', speaker['speech'].lower())),
+                len(speaker['speech'].split())),
+                sum(frequency for token in speech if token.pos_ == 'VERB'),
+                sum(frequency for token in speech if token.pos == 'NOUN')
             ])
             features.append(feat)
-            labels.append(0 if i % 2 == 0 else 1) # Assumes alternating roles
-
+            labels.append((0 if i % 2 == 0 else 1))
         scaler = StandardScaler()
         X = scaler.fit_transform(features)
-        
-        clf = RandomForestClassifier(n_estimators=150, max_depth=10, random_state=42, class_weight='balanced')
+        clf = RandomForestClassifier(
+            n_estimators=150, max_depth=10, random_state=42, class_weight='balanced'
+        )
         clf.fit(X, labels)
-        
-        logger.info("Saving trained models to disk...")
         joblib.dump(clf, os.path.join(OUTPUT_DIR, 'role_classifier.pkl'))
         joblib.dump(vectorizer, os.path.join(OUTPUT_DIR, 'text_vectorizer.pkl'))
         joblib.dump(scaler, os.path.join(OUTPUT_DIR, 'feature_scaler.pkl'))
-        
         return clf, vectorizer, scaler
     except Exception as e:
         logger.error(f"Classifier training failed: {str(e)}")
         raise
 
-def classify_roles(utterances: List[Dict], clf, vectorizer, scaler):
-    """Classifies roles using the pre-trained RandomForest model."""
+def classify_roles(speakers: List[Dict], clf, vectorizer, scaler):
     try:
-        logger.info("Classifying roles using trained model...")
-        texts = [u['text'] for u in utterances]
-        X_text = vectorizer.transform(texts)
+        speech = [u['speech'] for u in speakers]
+        X_text = vectorizer.transform(speech)
         results = []
-        for i, utterance in enumerate(utterances):
-            prosodic = utterance.get('prosodic_features', {})
+        for i, speaker in enumerate(speakers):
+            prosodic = speaker['speech_features']
             feat = [
-                prosodic.get('duration', 0), prosodic.get('mean_pitch', 0), prosodic.get('pitch_sd', 0),
-                prosodic.get('intensityMean', 0), prosodic.get('intensitySD', 0)
+                prosodic['duration'], prosodic['mean_pitch'], prosodic['min_pitch'],
+                prosodic['max_pitch'], prosodic['pitch_sd'], prosodic['intensityMean'],
+                prosodic['intensityMin'], prosodic['intensityMax'], prosodic['intensitySD'],
             ]
             feat.extend(X_text[i].toarray()[0].tolist())
-            doc = nlp(utterance['text'])
+            doc = nlp(speaker['speech'])
             feat.extend([
-                int(utterance['text'].endswith('?')),
-                len(re.findall(r'\b(why|how|what|when|where|who|which)\b', utterance['text'].lower())),
-                len(utterance['text'].split()),
+                int(speaker['speech'].endswith('?')),
+                len(re.findall(r'\b(why|how|what|when|where|who|which)\b', speaker['speech'].lower())),
+                len(speaker['speech'].split()),
                 sum(1 for token in doc if token.pos_ == 'VERB'),
                 sum(1 for token in doc if token.pos_ == 'NOUN')
             ])
             X = scaler.transform([feat])
             role = 'Interviewer' if clf.predict(X)[0] == 0 else 'Interviewee'
-            results.append({**utterance, 'role': role})
+            results.append({**speaker, 'role': role})
         return results
     except Exception as e:
-        logger.error(f"Role classification execution failed: {str(e)}")
+        logger.error(f"Role classification failed: {str(e)}")
         raise
 
-# --- 6. Analysis and Reporting Functions ---
-
-def analyze_interviewee_voice(audio_path: str, utterances: List[Dict]) -> Dict:
-    """Analyzes voice characteristics of all utterances classified as 'Interviewee'."""
+def analyze_interviewee_voice(audio_path: str, speakers: List[Dict]) -> Dict:
     try:
-        interviewee_utterances = [u for u in utterances if u.get('role') == 'Interviewee']
-        if not interviewee_utterances:
-            logger.warning("No interviewee utterances found to analyze.")
-            return {'error': 'No interviewee utterances found'}
-
-        logger.info(f"Analyzing {len(interviewee_utterances)} interviewee utterances...")
         y, sr = librosa.load(audio_path, sr=16000)
-        
-        segments = [y[int(u['start']*sr/1000):int(u['end']*sr/1000)] for u in interviewee_utterances]
-        
-        total_duration = sum(u['prosodic_features'].get('duration', 0) for u in interviewee_utterances)
-        total_words = sum(len(u['text'].split()) for u in interviewee_utterances)
-        speaking_rate = (total_words / total_duration) * 60 if total_duration > 0 else 0
-
-        filler_words = {'um', 'uh', 'like', 'you know', 'so', 'i mean', 'actually'}
-        filler_count = sum(1 for u in interviewee_utterances for word in u['text'].lower().split() if word in filler_words)
+        interviewee_speakers = [u for u in speakers if u['role'] == 'Interviewee']
+        if not interviewee_speakers:
+            return {'error': 'No interviewee speeches found'}
+        segments = []
+        for u in interviewee_speakers:
+            start = int(u['start'] * sr / 1000)
+            end = int(u['end'] * sr / 1000)
+            segments.append(y[start:end])
+        total_duration = sum(u['speech_features']['duration'] for u in interviewee_speakers)
+        total_words = sum(len(u['speech'].split()) for u in interviewee_speakers)
+        speaking_rate = total_words / total_duration if total_duration > 0 else 0
+        filler_words = ['um', 'uh', 'like', 'you know', 'so', 'i mean']
+        filler_count = sum(sum(u['speech'].lower().count(fw) for fw in filler_words) for u in interviewee_speakers)
         filler_ratio = filler_count / total_words if total_words > 0 else 0
-        
-        pitches = np.concatenate([librosa.pyin(s, fmin=librosa.note_to_hz('C2'), fmax=librosa.note_to_hz('C7'))[0] for s in segments if len(s)>0])
-        pitches = pitches[~np.isnan(pitches)]
-        
-        intensities = np.concatenate([librosa.feature.rms(y=s)[0] for s in segments if len(s)>0])
-
+        all_words = ' '.join(u['speech'].lower() for u in interviewee_speakers).split()
+        word_counts = {}
+        for i in range(len(all_words) - 1):
+            bigram = (all_words[i], all_words[i + 1])
+            word_counts[bigram] = word_counts.get(bigram, 0) + 1
+        repetition_score = sum(1 for count in word_counts.values() if count > 1) / len(word_counts) if word_counts else 0
+        pitches = []
+        for segment in segments:
+            f0, voiced_flag, _ = librosa.pyin(segment, fmin=80, fmax=300, sr=sr)
+            pitches.extend(f0[voiced_flag])
+        pitch_mean = np.mean(pitches) if len(pitches) > 0 else 0
         pitch_std = np.std(pitches) if len(pitches) > 0 else 0
-        intensity_std = np.std(intensities) if len(intensities) > 0 else 0
-
-        anxiety_score = max(0, min(1, pitch_std / 50))
-        confidence_score = max(0, min(1, 1 - (intensity_std * 10)))
-        hesitation_score = max(0, min(1, (filler_ratio * 2) + (pitch_std / 100)))
-
+        jitter = np.mean(np.abs(np.diff(pitches))) / pitch_mean if len(pitches) > 1 and pitch_mean > 0 else 0
+        intensities = []
+        for segment in segments:
+            rms = librosa.feature.rms(y=segment)[0]
+            intensities.extend(rms)
+        intensity_mean = np.mean(intensities) if intensities else 0
+        intensity_std = np.std(intensities) if intensities else 0
+        shimmer = np.mean(np.abs(np.diff(intensities))) / intensity_mean if len(intensities) > 1 and intensity_mean > 0 else 0
+        anxiety_score = 0.6 * (pitch_std / pitch_mean) + 0.4 * (jitter + shimmer) if pitch_mean > 0 else 0
+        confidence_score = 0.7 * (1 / (1 + intensity_std)) + 0.3 * (1 / (1 + filler_ratio))
+        hesitation_score = filler_ratio + repetition_score
+        anxiety_level = 'High' if anxiety_score > 0.15 else 'Moderate' if anxiety_score > 0.07 else 'Low'
+        confidence_level = 'High' if confidence_score > 0.7 else 'Moderate' if confidence_score > 0.5 else 'Low'
+        fluency_level = 'Fluent' if (filler_ratio < 0.05 and repetition_score < 0.1) else 'Moderate' if (filler_ratio < 0.1 and repetition_score < 0.2) else 'Disfluent'
         return {
             'speaking_rate': float(round(speaking_rate, 2)),
             'filler_ratio': float(round(filler_ratio, 4)),
-            'pitch_std_dev': float(round(pitch_std, 2)),
-            'intensity_std_dev': float(round(intensity_std, 4)),
-            'composite_scores': {
-                'anxiety': float(round(anxiety_score, 4)),
-                'confidence': float(round(confidence_score, 4)),
-                'hesitation': float(round(hesitation_score, 4))
-            }
+            'repetition_score': float(round(repetition_score, 4)),
+            'pitch_analysis': {'mean': float(round(pitch_mean, 2)), 'std_dev': float(round(pitch_std, 2)), 'jitter': float(round(jitter, 4))},
+            'intensity_analysis': {'mean': float(round(intensity_mean, 2)), 'std_dev': float(round(intensity_std, 2)), 'shimmer': float(round(shimmer, 4))},
+            'composite_scores': {'anxiety': float(round(anxiety_score, 4)), 'confidence': float(round(confidence_score, 4)), 'hesitation': float(round(hesitation_score, 4))},
+            'interpretation': {'anxiety_level': anxiety_level, 'confidence_level': confidence_level, 'fluency_level': fluency_level}
         }
     except Exception as e:
-        logger.error(f"Voice analysis failed: {str(e)}", exc_info=True)
+        logger.error(f"Voice analysis failed: {str(e)}")
         return {'error': str(e)}
 
+def generate_voice_interpretation(analysis: Dict) -> str:
+    if 'error' in analysis:
+        return "Voice analysis unavailable due to processing limitations."
+    interpretation_lines = [
+        "Vocal Performance Profile:",
+        f"- Speaking Rate: {analysis['speaking_rate']} words/sec - Benchmark: 2.0-3.0 wps for clear delivery",
+        f"- Filler Word Frequency: {analysis['filler_ratio'] * 100:.1f}% - Measures non-content words",
+        f"- Repetition Index: {analysis['repetition_score']:.3f} - Frequency of repeated phrases",
+        f"- Anxiety Indicator: {analysis['interpretation']['anxiety_level']} (Score: {analysis['composite_scores']['anxiety']:.3f}) - Pitch and vocal stability",
+        f"- Confidence Indicator: {analysis['interpretation']['confidence_level']} (Score: {analysis['composite_scores']['confidence']:.3f}) - Vocal strength",
+        f"- Fluency Rating: {analysis['interpretation']['fluency_level']} - Speech flow and coherence",
+        "",
+        "HR Insights:",
+        "- Rapid speech (>3.0 wps) may signal enthusiasm but risks clarity.",
+        "- High filler word use reduces perceived professionalism.",
+        "- Elevated anxiety suggests pressure; training can build resilience.",
+        "- Strong confidence aligns with leadership presence.",
+        "- Fluent speech enhances engagement, critical for team roles."
+    ]
+    return "\n".join(interpretation_lines)
+
+def generate_anxiety_confidence_chart(composite_scores: Dict, chart_path_or_buffer):
+    try:
+        labels = ['Anxiety', 'Confidence']
+        scores = [composite_scores.get('anxiety', 0), composite_scores.get('confidence', 0)]
+        fig, ax = plt.subplots(figsize=(5, 3.5))
+        bars = ax.bar(labels, scores, color=['#FF5252', '#26A69A'], edgecolor='black', width=0.45)
+        ax.set_ylabel('Score (Normalized)', fontsize=12)
+        ax.set_title('Vocal Dynamics: Anxiety vs. Confidence', fontsize=14, pad=15)
+        ax.set_ylim(0, 1.3)
+        for bar in bars:
+            height = bar.get_height()
+            ax.text(bar.get_x() + bar.get_width()/2, height + 0.05, f"{height:.2f}",
+                    ha='center', color='black', fontweight='bold', fontsize=11)
+        ax.grid(True, axis='y', linestyle='--', alpha=0.7)
+        plt.tight_layout()
+        plt.savefig(chart_path_or_buffer, format='png', bbox_inches='tight', dpi=300)
+        plt.close(fig)
+    except Exception as e:
+        logger.error(f"Error generating chart: {str(e)}")
+
+def calculate_acceptance_probability(analysis_data: Dict) -> float:
+    voice = analysis_data.get('voice_analysis', {})
+    if 'error' in voice: return 0.0
+    w_confidence, w_anxiety, w_fluency, w_speaking_rate, w_filler_repetition, w_content_strengths = 0.35, -0.25, 0.2, 0.15, -0.15, 0.25
+    confidence_score = voice.get('composite_scores', {}).get('confidence', 0.0)
+    anxiety_score = voice.get('composite_scores', {}).get('anxiety', 0.0)
+    fluency_level = voice.get('interpretation', {}).get('fluency_level', 'Disfluent')
+    speaking_rate = voice.get('speaking_rate', 0.0)
+    filler_ratio = voice.get('filler_ratio', 0.0)
+    repetition_score = voice.get('repetition_score', 0.0)
+    fluency_map = {'Fluent': 1.0, 'Moderate': 0.6, 'Disfluent': 0.2}
+    fluency_val = fluency_map.get(fluency_level, 0.2)
+    ideal_speaking_rate = 2.5
+    speaking_rate_deviation = abs(speaking_rate - ideal_speaking_rate)
+    speaking_rate_score = max(0, 1 - (speaking_rate_deviation / ideal_speaking_rate))
+    filler_repetition_composite = (filler_ratio + repetition_score) / 2
+    filler_repetition_score = max(0, 1 - filler_repetition_composite)
+    content_strength_val = 0.85 if analysis_data.get('text_analysis', {}).get('total_duration', 0) > 60 else 0.4
+    raw_score = (confidence_score * w_confidence + (1 - anxiety_score) * abs(w_anxiety) + fluency_val * w_fluency + speaking_rate_score * w_speaking_rate + filler_repetition_score * abs(w_filler_repetition) + content_strength_val * w_content_strengths)
+    max_possible_score = (w_confidence + abs(w_anxiety) + w_fluency + w_speaking_rate + abs(w_filler_repetition) + w_content_strengths)
+    if max_possible_score == 0: return 50.0
+    normalized_score = raw_score / max_possible_score
+    acceptance_probability = max(0.0, min(1.0, normalized_score))
+    return float(f"{acceptance_probability * 100:.2f}")
+
 def generate_report(analysis_data: Dict) -> str:
-    """Generates a text-based summary report using Gemini AI."""
     try:
-        logger.info("Generating final report text with Gemini...")
-        # ... (Your generate_report function logic here)
-        return "Gemini report text would be generated here."
+        voice = analysis_data.get('voice_analysis', {})
+        voice_interpretation = generate_voice_interpretation(voice)
+        interviewee_responses = [f"Speaker {u['speaker']} ({u['role']}): {u['text']}" for u in analysis_data['transcript'] if u['role'] == 'Interviewee'][:6]
+        acceptance_prob = analysis_data.get('acceptance_probability', None)
+        acceptance_line = ""
+        if acceptance_prob is not None:
+            acceptance_line = f"\n**Hiring Suitability Score: {acceptance_prob:.2f}%**\n"
+            if acceptance_prob >= 80: acceptance_line += "HR Verdict: Outstanding candidate, highly recommended for immediate advancement."
+            elif acceptance_prob >= 60: acceptance_line += "HR Verdict: Strong candidate, suitable for further evaluation with targeted development."
+            elif acceptance_prob >= 40: acceptance_line += "HR Verdict: Moderate potential, requires additional assessment and skill-building."
+            else: acceptance_line += "HR Verdict: Limited fit, significant improvement needed for role alignment."
+        prompt = f"""
+        You are EvalBot, a senior HR consultant with 20+ years of experience, delivering a polished, concise, and engaging interview analysis report. Use a professional tone, clear headings, and bullet points ('- ') for readability. Avoid redundancy and ensure distinct sections for strengths, growth areas, and recommendations.
+        {acceptance_line}
+        **1. Executive Summary**
+        - Provide a concise overview of performance, key metrics, and hiring potential.
+        - Interview length: {analysis_data['text_analysis']['total_duration']:.2f} seconds
+        - Speaker turns: {analysis_data['text_analysis']['speaker_turns']}
+        - Participants: {', '.join(analysis_data['speakers'])}
+        **2. Communication and Vocal Dynamics**
+        - Evaluate vocal delivery (rate, fluency, confidence) and professional impact.
+        - Offer HR insights on workplace alignment.
+        {voice_interpretation}
+        **3. Competency and Content Evaluation**
+        - Assess competencies: leadership, problem-solving, communication, adaptability.
+        - List strengths and growth areas separately, with specific examples.
+        - Sample responses:
+        {chr(10).join(interviewee_responses)}
+        **4. Role Fit and Growth Potential**
+        - Analyze cultural fit, role readiness, and long-term potential.
+        - Highlight enthusiasm and scalability.
+        **5. Strategic HR Recommendations**
+        - Provide distinct, prioritized strategies for candidate growth.
+        - Target: Communication, Response Depth, Professional Presence.
+        - List clear next steps for hiring managers (e.g., advance, train, assess).
+        """
+        response = gemini_model.generate_content(prompt)
+        return response.text
     except Exception as e:
         logger.error(f"Report generation failed: {str(e)}")
-        return f"Error in report generation: {str(e)}"
+        return f"Error generating report: {str(e)}"
+
+def create_pdf_report(analysis_data: Dict, output_path: str, gemini_report_text: str):
+    try:
+        doc = SimpleDocTemplate(output_path, pagesize=letter,
+                                rightMargin=0.7*inch, leftMargin=0.7*inch,
+                                topMargin=0.9*inch, bottomMargin=0.9*inch)
+        styles = getSampleStyleSheet()
+        h1 = ParagraphStyle(name='Heading1', fontSize=22, leading=26, spaceAfter=20, alignment=1, textColor=colors.HexColor('#003087'), fontName='Helvetica-Bold')
+        h2 = ParagraphStyle(name='Heading2', fontSize=15, leading=18, spaceBefore=14, spaceAfter=8, textColor=colors.HexColor('#0050BC'), fontName='Helvetica-Bold')
+        h3 = ParagraphStyle(name='Heading3', fontSize=11, leading=14, spaceBefore=10, spaceAfter=6, textColor=colors.HexColor('#3F7CFF'), fontName='Helvetica')
+        body_text = ParagraphStyle(name='BodyText', fontSize=10, leading=13, spaceAfter=8, fontName='Helvetica', textColor=colors.HexColor('#333333'))
+        bullet_style = ParagraphStyle(name='Bullet', parent=body_text, leftIndent=20, bulletIndent=10, fontName='Helvetica', bulletFontName='Helvetica', bulletFontSize=10)
+        
+        story = []
+
+        def header_footer(canvas, doc):
+            canvas.saveState()
+            canvas.setFont('Helvetica', 8)
+            canvas.setFillColor(colors.HexColor('#666666'))
+            canvas.drawString(doc.leftMargin, 0.4 * inch, f"Page {doc.page} | EvalBot HR Interview Report | Confidential")
+            canvas.setStrokeColor(colors.HexColor('#0050BC'))
+            canvas.setLineWidth(1)
+            canvas.line(doc.leftMargin, doc.height + 0.85*inch, doc.width + doc.leftMargin, doc.height + 0.85*inch)
+            canvas.setFont('Helvetica-Bold', 10)
+            canvas.drawString(doc.leftMargin, doc.height + 0.9*inch, "Candidate Interview Analysis")
+            canvas.drawRightString(doc.width + doc.leftMargin, doc.height + 0.9*inch, time.strftime('%B %d, %Y'))
+            canvas.restoreState()
+
+        # Title Page
+        story.append(Paragraph("Candidate Interview Analysis", h1))
+        story.append(Paragraph(f"Generated: {time.strftime('%B %d, %Y')}", ParagraphStyle(name='Date', alignment=1, fontSize=10, textColor=colors.HexColor('#666666'), fontName='Helvetica')))
+        story.append(Spacer(1, 0.5 * inch))
+        acceptance_prob = analysis_data.get('acceptance_probability')
+        if acceptance_prob is not None:
+            story.append(Paragraph("Hiring Suitability Snapshot", h2))
+            prob_color = colors.HexColor('#2E7D32') if acceptance_prob >= 80 else (colors.HexColor('#F57C00') if acceptance_prob >= 60 else colors.HexColor('#D32F2F'))
+            story.append(Paragraph(f"Suitability Score: <font size=16 color='{prob_color.hexval()}'><b>{acceptance_prob:.2f}%</b></font>",
+                                 ParagraphStyle(name='Prob', fontSize=12, spaceAfter=12, alignment=1, fontName='Helvetica-Bold')))
+            if acceptance_prob >= 80:
+                story.append(Paragraph("<b>HR Verdict:</b> Outstanding candidate, highly recommended for immediate advancement.", body_text))
+            elif acceptance_prob >= 60:
+                story.append(Paragraph("<b>HR Verdict:</b> Strong candidate, suitable for further evaluation with targeted development.", body_text))
+            elif acceptance_prob >= 40:
+                story.append(Paragraph("<b>HR Verdict:</b> Moderate potential, requires additional assessment and skill-building.", body_text))
+            else:
+                story.append(Paragraph("<b>HR Verdict:</b> Limited fit, significant improvement needed for role alignment.", body_text))
+            story.append(Spacer(1, 0.3 * inch))
+            table_data = [
+                ['Metric', 'Value'],
+                ['Interview Duration', f"{analysis_data['text_analysis']['total_duration']:.2f} seconds"],
+                ['Speaker Turns', f"{analysis_data['text_analysis']['speaker_turns']}"],
+                ['Participants', ', '.join(sorted(analysis_data['speakers']))]
+            ]
+            table = Table(table_data, colWidths=[2.2*inch, 3.8*inch])
+            table.setStyle(TableStyle([
+                ('BACKGROUND', (0,0), (-1,0), colors.HexColor('#0050BC')),
+                ('TEXTCOLOR', (0,0), (-1,0), colors.white),
+                ('ALIGN', (0,0), (-1,-1), 'LEFT'),
+                ('VALIGN', (0,0), (-1,-1), 'MIDDLE'),
+                ('FONTNAME', (0,0), (-1,0), 'Helvetica-Bold'),
+                ('FONTSIZE', (0,0), (-1,-1), 9),
+                ('BOTTOMPADDING', (0,0), (-1,0), 10),
+                ('TOPPADDING', (0,0), (-1,0), 10),
+                ('BACKGROUND', (0,1), (-1,-1), colors.HexColor('#F5F6FA')),
+                ('GRID', (0,0), (-1,-1), 0.5, colors.HexColor('#DDE4EB'))
+            ]))
+            story.append(table)
+        story.append(Spacer(1, 0.4 * inch))
+        story.append(Paragraph("Prepared by: EvalBot - AI-Powered HR Analysis", body_text))
+        story.append(PageBreak())
+
+        # Detailed Analysis
+        story.append(Paragraph("Detailed Candidate Evaluation", h1))
+        
+        # Communication and Vocal Dynamics
+        story.append(Paragraph("1. Communication & Vocal Dynamics", h2))
+        voice_analysis = analysis_data.get('voice_analysis', {})
+        if voice_analysis and 'error' not in voice_analysis:
+            table_data = [
+                ['Metric', 'Value', 'HR Insight'],
+                ['Speaking Rate', f"{voice_analysis.get('speaking_rate', 0):.2f} words/sec", 'Benchmark: 2.0-3.0 wps; impacts clarity'],
+                ['Filler Words', f"{voice_analysis.get('filler_ratio', 0) * 100:.1f}%", 'High usage reduces credibility'],
+                ['Anxiety', voice_analysis.get('interpretation', {}).get('anxiety_level', 'N/A'), f"Score: {voice_analysis.get('composite_scores', {}).get('anxiety', 0):.3f}; stress response"],
+                ['Confidence', voice_analysis.get('interpretation', {}).get('confidence_level', 'N/A'), f"Score: {voice_analysis.get('composite_scores', {}).get('confidence', 0):.3f}; vocal strength"],
+                ['Fluency', voice_analysis.get('interpretation', {}).get('fluency_level', 'N/A'), 'Drives engagement']
+            ]
+            table = Table(table_data, colWidths=[1.7*inch, 1.2*inch, 3.1*inch])
+            table.setStyle(TableStyle([
+                ('BACKGROUND', (0,0), (-1,0), colors.HexColor('#0050BC')),
+                ('TEXTCOLOR', (0,0), (-1,0), colors.white),
+                ('ALIGN', (0,0), (-1,-1), 'LEFT'),
+                ('VALIGN', (0,0), (-1,-1), 'MIDDLE'),
+                ('FONTNAME', (0,0), (-1,0), 'Helvetica-Bold'),
+                ('FONTSIZE', (0,0), (-1,-1), 9),
+                ('BOTTOMPADDING', (0,0), (-1,0), 10),
+                ('TOPPADDING', (0,0), (-1,0), 10),
+                ('BACKGROUND', (0,1), (-1,-1), colors.HexColor('#F5F6FA')),
+                ('GRID', (0,0), (-1,-1), 0.5, colors.HexColor('#DDE4EB'))
+            ]))
+            story.append(table)
+            story.append(Spacer(1, 0.2 * inch))
+            chart_buffer = io.BytesIO()
+            generate_anxiety_confidence_chart(voice_analysis.get('composite_scores', {}), chart_buffer)
+            chart_buffer.seek(0)
+            img = Image(chart_buffer, width=4.8*inch, height=3.2*inch)
+            img.hAlign = 'CENTER'
+            story.append(img)
+        else:
+            story.append(Paragraph("Vocal analysis unavailable.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+
+        # Parse Gemini Report
+        sections = {
+            "Executive Summary": [],
+            "Communication and Vocal Dynamics": [],
+            "Competency and Content Evaluation": {"Strengths": [], "Growth Areas": []},
+            "Role Fit and Growth Potential": [],
+            "Strategic HR Recommendations": {"Development Priorities": [], "Next Steps": []}
+        }
+        report_parts = re.split(r'(\s*\*\*\s*\d\.\s*.*?\s*\*\*)', gemini_report_text)
+        current_section = None
+        for part in report_parts:
+            if not part.strip(): continue
+            is_heading = False
+            for title in sections.keys():
+                if title.lower() in part.lower():
+                    current_section = title
+                    is_heading = True
+                    break
+            if not is_heading and current_section:
+                if current_section == "Competency and Content Evaluation":
+                    if 'strength' in part.lower() or any(k in part.lower() for k in ['leadership', 'problem-solving', 'communication', 'adaptability']):
+                        sections[current_section]["Strengths"].append(part.strip())
+                    elif 'improve' in part.lower() or 'grow' in part.lower() or 'challenge' in part.lower():
+                        sections[current_section]["Growth Areas"].append(part.strip())
+                elif current_section == "Strategic HR Recommendations":
+                    if any(k in part.lower() for k in ['communication', 'depth', 'presence', 'improve']):
+                        sections[current_section]["Development Priorities"].append(part.strip())
+                    elif any(k in part.lower() for k in ['advance', 'train', 'assess', 'next step']):
+                        sections[current_section]["Next Steps"].append(part.strip())
+                else:
+                    sections[current_section].append(part.strip())
+
+        # Executive Summary
+        story.append(Paragraph("2. Executive Summary", h2))
+        if sections['Executive Summary']:
+            for line in sections['Executive Summary']:
+                if line.startswith(('-', '•', '*')):
+                    story.append(Paragraph(line.lstrip('-•* ').strip(), bullet_style))
+                else:
+                    story.append(Paragraph(line, body_text))
+        else:
+            story.append(Paragraph("Summary unavailable.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+
+        # Competency and Content
+        story.append(Paragraph("3. Competency & Content", h2))
+        story.append(Paragraph("Strengths", h3))
+        if sections['Competency and Content Evaluation']['Strengths']:
+            for line in sections['Competency and Content Evaluation']['Strengths']:
+                story.append(Paragraph(line.lstrip('-•* ').strip(), bullet_style))
+        else:
+            story.append(Paragraph("No strengths identified.", body_text))
+        story.append(Spacer(1, 0.2 * inch))
+        story.append(Paragraph("Growth Areas", h3))
+        if sections['Competency and Content Evaluation']['Growth Areas']:
+            for line in sections['Competency and Content Evaluation']['Growth Areas']:
+                story.append(Paragraph(line.lstrip('-•* ').strip(), bullet_style))
+        else:
+            story.append(Paragraph("No growth areas identified.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+
+        # Role Fit
+        story.append(Paragraph("4. Role Fit & Potential", h2))
+        if sections['Role Fit and Growth Potential']:
+            for line in sections['Role Fit and Growth Potential']:
+                if line.startswith(('-', '•', '*')):
+                    story.append(Paragraph(line.lstrip('-•* ').strip(), bullet_style))
+                else:
+                    story.append(Paragraph(line, body_text))
+        else:
+            story.append(Paragraph("Fit and potential analysis unavailable.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+
+        # Strategic Recommendations
+        story.append(Paragraph("5. Strategic Recommendations", h2))
+        story.append(Paragraph("Development Priorities", h3))
+        if sections['Strategic HR Recommendations']['Development Priorities']:
+            for line in sections['Strategic HR Recommendations']['Development Priorities']:
+                story.append(Paragraph(line.lstrip('-•* ').strip(), bullet_style))
+        else:
+            story.append(Paragraph("No development priorities specified.", body_text))
+        story.append(Spacer(1, 0.2 * inch))
+        story.append(Paragraph("Next Steps for Managers", h3))
+        if sections['Strategic HR Recommendations']['Next Steps']:
+            for line in sections['Strategic HR Recommendations']['Next Steps']:
+                story.append(Paragraph(line.lstrip('-•* ').strip(), bullet_style))
+        else:
+            story.append(Paragraph("No next steps provided.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+        story.append(Paragraph("This report provides a data-driven evaluation to guide hiring and development decisions.", body_text))
+
+        doc.build(story, onFirstPage=header_footer, onLaterPages=header_footer)
+        return True
+    except Exception as e:
+        logger.error(f"PDF creation failed: {str(e)}", exc_info=True)
+        return False
 
-# --- 7. Main Orchestration Function ---
+def convert_to_serializable(obj):
+    if isinstance(obj, np.generic): return obj.item()
+    if isinstance(obj, dict): return {k: convert_to_serializable(v) for k, v in obj.items()}
+    if isinstance(obj, list): return [convert_to_serializable(i) for i in obj]
+    if isinstance(obj, np.ndarray): return obj.tolist()
+    return obj
 
-def process_interview_from_url(audio_url: str):
-    """
-    Main pipeline to download, process, and analyze an interview from a URL.
-    """
+def process_interview(audio_path_or_url: str):
     local_audio_path = None
     wav_file = None
-    
+    is_downloaded = False
     try:
-        # Step 1: Download and Convert
-        local_audio_path = download_audio_to_temp_file(audio_url)
+        logger.info(f"Starting processing for {audio_path_or_url}")
+        if audio_path_or_url.startswith(('http://', 'https://')):
+            local_audio_path = download_audio_from_url(audio_path_or_url)
+            is_downloaded = True
+        else:
+            local_audio_path = audio_path_or_url
         wav_file = convert_to_wav(local_audio_path)
-        
-        # Step 2: Transcribe and Diarize
         transcript = transcribe(wav_file)
-        
-        # Step 3: Extract Features
-        logger.info("Extracting prosodic features for all utterances...")
-        with ThreadPoolExecutor() as executor:
-            futures = {executor.submit(extract_prosodic_features, wav_file, u['start'], u['end']): u for u in transcript['utterances']}
-            for future in futures:
-                utterance = futures[future]
-                utterance['prosodic_features'] = future.result()
-
-        # Step 4: Classify Roles
-        classifier_path = os.path.join(OUTPUT_DIR, 'role_classifier.pkl')
-        if os.path.exists(classifier_path):
-            logger.info("Loading existing role classifier model.")
-            clf = joblib.load(classifier_path)
+        for utterance in transcript['utterances']:
+            utterance['prosodic_features'] = extract_prosodic_features(wav_file, utterance['start'], utterance['end'])
+        utterances_with_speakers = identify_speakers(transcript, wav_file)
+        clf, vectorizer, scaler = None, None, None
+        if os.path.exists(os.path.join(OUTPUT_DIR, 'role_classifier.pkl')):
+            clf = joblib.load(os.path.join(OUTPUT_DIR, 'role_classifier.pkl'))
             vectorizer = joblib.load(os.path.join(OUTPUT_DIR, 'text_vectorizer.pkl'))
             scaler = joblib.load(os.path.join(OUTPUT_DIR, 'feature_scaler.pkl'))
         else:
-            clf, vectorizer, scaler = train_role_classifier(transcript['utterances'])
-            
-        classified_utterances = classify_roles(transcript['utterances'], clf, vectorizer, scaler)
-        
-        # Step 5: Analyze Voice and Generate Report
+            clf, vectorizer, scaler = train_role_classifier(utterances_with_speakers)
+        classified_utterances = classify_roles(utterances_with_speakers, clf, vectorizer, scaler)
         voice_analysis = analyze_interviewee_voice(wav_file, classified_utterances)
-        
         analysis_data = {
             'transcript': classified_utterances,
-            'speakers': list(set(u['speaker'] for u in classified_utterances if u.get('speaker'))),
+            'speakers': list(set(u['speaker'] for u in classified_utterances)),
             'voice_analysis': voice_analysis,
             'text_analysis': {
-                'total_duration': transcript.get('audio_duration', 0),
+                'total_duration': sum(u['prosodic_features']['duration'] for u in classified_utterances),
                 'speaker_turns': len(classified_utterances)
             }
         }
-        
+        analysis_data['acceptance_probability'] = calculate_acceptance_probability(analysis_data)
         gemini_report_text = generate_report(analysis_data)
-        
-        # Step 6: Save Results
         base_name = str(uuid.uuid4())
+        pdf_path = os.path.join(OUTPUT_DIR, f"{base_name}_report.pdf")
         json_path = os.path.join(OUTPUT_DIR, f"{base_name}_analysis.json")
-        
+        create_pdf_report(analysis_data, pdf_path, gemini_report_text=gemini_report_text)
         with open(json_path, 'w') as f:
-            # Use default=str to handle any non-serializable data types gracefully
-            json.dump(analysis_data, f, indent=4, default=str)
-
-        logger.info(f"Processing completed. Analysis saved to: {json_path}")
-        return {'json_path': json_path, 'report_text': gemini_report_text}
-        
+            serializable_data = convert_to_serializable(analysis_data)
+            json.dump(serializable_data, f, indent=2)
+        logger.info(f"Processing completed for {audio_path_or_url}")
+        return {'pdf_path': pdf_path, 'json_path': json_path}
     except Exception as e:
-        logger.error(f"Main processing pipeline failed for URL {audio_url}: {str(e)}", exc_info=True)
+        logger.error(f"Processing failed for {audio_path_or_url}: {str(e)}", exc_info=True)
         raise
-        
     finally:
-        # Step 7: Cleanup
-        logger.info("Cleaning up temporary files...")
         if wav_file and os.path.exists(wav_file):
-            try:
-                os.remove(wav_file)
-                logger.info(f"Removed temporary WAV file: {wav_file}")
-            except OSError as e:
-                logger.error(f"Error removing WAV file {wav_file}: {e}")
-        if local_audio_path and os.path.exists(local_audio_path):
-            try:
-                os.remove(local_audio_path)
-                logger.info(f"Removed temporary downloaded file: {local_audio_path}")
-            except OSError as e:
-                logger.error(f"Error removing downloaded file {local_audio_path}: {e}")
+            os.remove(wav_file)
+        if is_downloaded and local_audio_path and os.path.exists(local_audio_path):
+            os.remove(local_audio_path)
+            logger.info(f"Cleaned up temporary downloaded file: {local_audio_path}")