Update process_interview.py

process_interview.py

@@ -19,15 +19,16 @@ 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, PageBreak, Image, HRFlowable
+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, pipeline
+from transformers import AutoTokenizer, AutoModel
 import spacy
 import google.generativeai as genai
 import joblib
@@ -35,10 +36,12 @@ from concurrent.futures import ThreadPoolExecutor
 
 # Setup logging
 logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
+logger = logging.getLogger(_name_)
 logging.getLogger("nemo_logging").setLevel(logging.ERROR)
+logging.getLogger("nemo").setLevel(logging.ERROR)
 
 # Configuration
+AUDIO_DIR = "./uploads"
 OUTPUT_DIR = "./processed_audio"
 os.makedirs(OUTPUT_DIR, exist_ok=True)
 
@@ -47,21 +50,18 @@ PINECONE_KEY = os.getenv("PINECONE_KEY")
 ASSEMBLYAI_KEY = os.getenv("ASSEMBLYAI_KEY")
 GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")
 
-# --- All your original helper functions ---
-# I am including them exactly as you last provided them.
-# --- HELPER FUNCTION to download from URL ---
 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_filename = f"{uuid.uuid4()}.tmp_audio"
-        local_filename = os.path.join(temp_dir, temp_filename)
-        logger.info(f"Downloading audio from {url} to {local_filename}")
+        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(local_filename, 'wb') as f:
+            with open(temp_path, 'wb') as f:
                 for chunk in r.iter_content(chunk_size=8192):
                     f.write(chunk)
-        return local_filename
+        return temp_path
     except Exception as e:
         logger.error(f"Failed to download audio from URL {url}: {e}")
         raise
@@ -71,7 +71,12 @@ def initialize_services():
         pc = Pinecone(api_key=PINECONE_KEY)
         index_name = "interview-speaker-embeddings"
         if index_name not in pc.list_indexes().names():
-            pc.create_index(name=index_name, dimension=192, metric="cosine", spec=ServerlessSpec(cloud="aws", region="us-east-1"))
+            pc.create_index(
+                name=index_name,
+                dimension=192,
+                metric="cosine",
+                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')
@@ -87,8 +92,12 @@ 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=device)
+        model = EncDecSpeakerLabelModel.from_pretrained(
+            "nvidia/speakerverification_en_titanet_large",
+            map_location=torch.device('cpu')
+        )
         model.eval()
         return model
     except Exception as e:
@@ -108,7 +117,8 @@ speaker_model, nlp, tokenizer, llm_model = load_models()
 def convert_to_wav(audio_path: str, output_dir: str = OUTPUT_DIR) -> str:
     try:
         audio = AudioSegment.from_file(audio_path)
-        if audio.channels > 1: audio = audio.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")
@@ -121,14 +131,13 @@ def extract_prosodic_features(audio_path: str, start_ms: int, end_ms: int) -> Di
     try:
         audio = AudioSegment.from_file(audio_path)
         segment = audio[start_ms:end_ms]
-        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
-            segment.export(tmp.name, format="wav")
-            y, sr = librosa.load(tmp.name, sr=16000)
-            os.remove(tmp.name)
-        pitches, _ = librosa.piptrack(y=y, sr=sr)
+        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]
-        return {
-            'duration': (end_ms - start_ms) / 1000.0,
+        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,
@@ -138,132 +147,277 @@ def extract_prosodic_features(audio_path: str, start_ms: int, end_ms: int) -> Di
             '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 {}
+        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:
     try:
         with open(audio_path, 'rb') as f:
-            upload_response = requests.post("https://api.assemblyai.com/v2/upload", headers={"authorization": ASSEMBLYAI_KEY}, 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']
-        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_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']
         while True:
-            result = requests.get(f"https://api.assemblyai.com/v2/transcript/{transcript_id}", headers={"authorization": ASSEMBLYAI_KEY}).json()
-            if result['status'] == 'completed': return result
-            elif result['status'] == 'error': raise Exception(f"AssemblyAI Error: {result.get('error')}")
+            result = requests.get(
+                f"https://api.assemblyai.com/v2/transcript/{transcript_id}",
+                headers={"authorization": ASSEMBLYAI_KEY}
+            ).json()
+            if result['status'] == 'completed':
+                return result
+            elif result['status'] == 'error':
+                raise Exception(result['error'])
             time.sleep(5)
     except Exception as e:
         logger.error(f"Transcription failed: {str(e)}")
         raise
 
-def process_utterance(utterance, full_audio):
+def process_utterance(utterance, full_audio, wav_file):
     try:
-        start, end = utterance['start'], utterance['end']
+        start = utterance['start']
+        end = utterance['end']
         segment = full_audio[start:end]
-        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
-            segment.export(tmp.name, format="wav")
-            with torch.no_grad():
-                embedding = speaker_model.get_embedding(tmp.name).cpu().numpy()
-            os.remove(tmp.name)
+        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'] > 0.75:
+        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"speaker_{uuid.uuid4().hex[:6]}"
-            speaker_name = f"Speaker_{speaker_id[-4:].upper()}"
+            speaker_id = f"unknown_{uuid.uuid4().hex[:6]}"
+            speaker_name = f"Speaker_{speaker_id[-4:]}"
             index.upsert([(speaker_id, embedding_list, {"speaker_name": speaker_name})])
-        return {**utterance, 'speaker': speaker_name, 'speaker_id': speaker_id}
+        os.remove(temp_path)
+        return {
+            **utterance,
+            'speaker': speaker_name,
+            'speaker_id': speaker_id,
+            'embedding': embedding_list
+        }
     except Exception as e:
         logger.error(f"Utterance processing failed: {str(e)}", exc_info=True)
-        return {**utterance, 'speaker': 'Unknown', 'speaker_id': 'unknown'}
+        return {
+            **utterance,
+            'speaker': 'Unknown',
+            'speaker_id': 'unknown',
+            'embedding': None
+        }
 
 def identify_speakers(transcript: Dict, wav_file: str) -> List[Dict]:
     try:
         full_audio = AudioSegment.from_wav(wav_file)
+        utterances = transcript['utterances']
         with ThreadPoolExecutor(max_workers=5) as executor:
-            futures = [executor.submit(process_utterance, u, full_audio) for u in transcript['utterances']]
+            futures = [
+                executor.submit(process_utterance, utterance, full_audio, wav_file)
+                for utterance in utterances
+            ]
             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]):
+    try:
+        texts = [u['text'] for u in utterances]
+        vectorizer = TfidfVectorizer(max_features=500, ngram_range=(1, 2))
+        X_text = vectorizer.fit_transform(texts)
+        features = []
+        labels = []
+        for i, utterance in enumerate(utterances):
+            prosodic = utterance['prosodic_features']
+            feat = [
+                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'])
+            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')
+            ])
+            features.append(feat)
+            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.fit(X, labels)
+        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):
+    try:
+        texts = [u['text'] for u in utterances]
+        X_text = vectorizer.transform(texts)
+        results = []
+        for i, utterance in enumerate(utterances):
+            prosodic = utterance['prosodic_features']
+            feat = [
+                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'])
+            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')
+            ])
+            X = scaler.transform([feat])
+            role = 'Interviewer' if clf.predict(X)[0] == 0 else 'Interviewee'
+            results.append({**utterance, 'role': role})
+        return results
+    except Exception as e:
+        logger.error(f"Role classification failed: {str(e)}")
+        raise
+
 def analyze_interviewee_voice(audio_path: str, utterances: List[Dict]) -> Dict:
     try:
         y, sr = librosa.load(audio_path, sr=16000)
-        interviewee_utterances = [u for u in utterances if u.get('role') == 'Interviewee']
-        if not interviewee_utterances: return {'error': 'No interviewee utterances found'}
-        segments = [y[int(u['start']*sr/1000):int(u['end']*sr/1000)] for u in interviewee_utterances]
+        interviewee_utterances = [u for u in utterances if u['role'] == 'Interviewee']
+        if not interviewee_utterances:
+            return {'error': 'No interviewee utterances found'}
+        segments = []
+        for u in interviewee_utterances:
+            start = int(u['start'] * sr / 1000)
+            end = int(u['end'] * sr / 1000)
+            segments.append(y[start:end])
         total_duration = sum(u['prosodic_features']['duration'] for u in interviewee_utterances)
         total_words = sum(len(u['text'].split()) for u in interviewee_utterances)
         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['text'].lower().count(fw) for fw in filler_words) for u in interviewee_utterances)
         filler_ratio = filler_count / total_words if total_words > 0 else 0
-        repetition_score = 0
-        pitches, intensities = [], []
+        all_words = ' '.join(u['text'].lower() for u in interviewee_utterances).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:
-            if len(segment) == 0: continue
-            f0, voiced_flag, _ = librosa.pyin(segment, fmin=librosa.note_to_hz('C2'), fmax=librosa.note_to_hz('C7'), sr=sr)
+            f0, voiced_flag, _ = librosa.pyin(segment, fmin=80, fmax=300, sr=sr)
             pitches.extend(f0[voiced_flag])
-            intensities.extend(librosa.feature.rms(y=segment)[0])
         pitch_mean = np.mean(pitches) if len(pitches) > 0 else 0
-        intensity_std = np.std(intensities) if len(intensities) > 0 else 0
+        pitch_std = np.std(pitches) if len(pitches) > 0 else 0
         jitter = np.mean(np.abs(np.diff(pitches))) / pitch_mean if len(pitches) > 1 and pitch_mean > 0 else 0
-        shimmer = np.mean(np.abs(np.diff(intensities))) / np.mean(intensities) if len(intensities) > 1 and np.mean(intensities) > 0 else 0
-        anxiety_score = (0.6 * (np.std(pitches)/pitch_mean if pitch_mean > 0 else 0) + 0.4 * (jitter + shimmer))
-        confidence_score = 0.7 * (1/(1+intensity_std)) + 0.3 * (1/(1+filler_ratio))
+        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)), 'repetition_score': float(round(repetition_score, 4)),
+            'speaking_rate': float(round(speaking_rate, 2)),
+            'filler_ratio': float(round(filler_ratio, 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': '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 'disfluent'
-            }
+            '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)}")
         return {'error': str(e)}
 
-
-def analyze_text_content(utterances: List[Dict]) -> Dict:
-    interviewee_utterances = [u['text'] for u in utterances if u.get('role') == 'Interviewee']
-    if not interviewee_utterances:
-        return {"overall_sentiment": {"label": "NEUTRAL", "score": 1.0}, "mentioned_technologies": [], "mentioned_soft_skills": []}
-    
-    full_text = " ".join(interviewee_utterances)
-    sentiment_results = sentiment_pipeline(full_text, truncation=True, max_length=512)
-    
-    tech_keywords = ['python', 'react', 'aws', 'docker', 'api', 'fastapi', 'machine learning', 'pytorch', 'tensorflow']
-    soft_skills = ['leadership', 'teamwork', 'communication', 'problem solving', 'management', 'planning']
-    
-    found_tech = [kw for kw in tech_keywords if kw in full_text.lower()]
-    found_skills = [skill for skill in soft_skills if skill in full_text.lower()]
-    
-    return {
-        "overall_sentiment": sentiment_results[0],
-        "mentioned_technologies": list(set(found_tech)),
-        "mentioned_soft_skills": list(set(found_skills))
-    }
+def generate_voice_interpretation(analysis: Dict) -> str:
+    if 'error' in analysis:
+        return "Voice analysis not available due to processing error."
+    interpretation_lines = [
+        "Voice and Speech Profile:",
+        f"- Speaking Rate: {analysis['speaking_rate']} words/sec - Compared to optimal range (2.0-3.0 words/sec)",
+        f"- Filler Word Usage: {analysis['filler_ratio'] * 100:.1f}% - Frequency of non-content words (e.g., 'um', 'like')",
+        f"- Repetition Tendency: {analysis['repetition_score']:.3f} - Measure of repeated phrases",
+        f"- Anxiety Indicator: {analysis['interpretation']['anxiety_level']} (Score: {analysis['composite_scores']['anxiety']:.3f}) - Based on pitch and voice stability",
+        f"- Confidence Indicator: {analysis['interpretation']['confidence_level']} (Score: {analysis['composite_scores']['confidence']:.3f}) - Derived from vocal consistency",
+        f"- Fluency Assessment: {analysis['interpretation']['fluency_level']} - Reflects speech flow and coherence",
+        "",
+        "HR Insights:",
+        "- Faster speaking rates may indicate confidence but can suggest nervousness if excessive.",
+        "- High filler word usage often reduces perceived professionalism and clarity.",
+        "- Elevated anxiety indicators (pitch variability, jitter) may reflect interview pressure.",
+        "- Strong confidence scores suggest effective vocal presence and control.",
+        "- Fluency impacts listener engagement; disfluency may hinder communication effectiveness."
+    ]
+    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=(4, 2.5))
+        bars = ax.bar(labels, scores, color=['#FF6B6B', '#4ECDC4'], edgecolor='black')
+        ax.set_ylabel('Score (Normalized)')
+        ax.set_title('Vocal Dynamics: Anxiety vs. Confidence')
+        ax.set_ylim(0, 1.2)
+        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=10)
+        plt.tight_layout()
+        plt.savefig(chart_path_or_buffer, format='png', bbox_inches='tight', dpi=150)
+        plt.close(fig)
+    except Exception as e:
+        logger.error(f"Error generating chart: {str(e)}")
 
 def calculate_acceptance_probability(analysis_data: Dict) -> float:
-    # Your full, detailed function
     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 = 0.4, -0.3, 0.2, 0.1, -0.1, 0.2
     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')
+    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': 0.5, 'disfluent': 0.0}
+    fluency_map = {'Fluent': 1.0, 'Moderate': 0.5, 'Disfluent': 0.0}
     fluency_val = fluency_map.get(fluency_level, 0.0)
     ideal_speaking_rate = 2.5
     speaking_rate_deviation = abs(speaking_rate - ideal_speaking_rate)
@@ -278,55 +432,42 @@ def calculate_acceptance_probability(analysis_data: Dict) -> float:
     acceptance_probability = max(0.0, min(1.0, normalized_score))
     return float(f"{acceptance_probability * 100:.2f}")
 
-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
-
-# --- NEW: HR Persona Report Generation ---
 def generate_report(analysis_data: Dict) -> str:
     try:
         voice = analysis_data.get('voice_analysis', {})
-        voice_interp = "Voice analysis data was not available."
-        if voice and 'error' not in voice:
-            voice_interp = (f"The candidate's voice profile indicates a '{voice.get('interpretation', {}).get('confidence_level', 'N/A').upper()}' confidence level "
-                          f"and a '{voice.get('interpretation', {}).get('anxiety_level', 'N/A').upper()}' anxiety level. "
-                          f"Fluency was rated as '{voice.get('interpretation', {}).get('fluency_level', 'N/A').upper()}'.")
-
-        content = analysis_data.get('advanced_content_analysis', {})
-        content_interp = (f"Sentiment of responses was generally '{content.get('overall_sentiment', {}).get('label', 'N/A')}'. "
-                        f"Mentioned technical skills: {', '.join(content.get('mentioned_technologies', [])) or 'None'}. "
-                        f"Mentioned soft skills: {', '.join(content.get('mentioned_soft_skills', [])) or 'None'}.")
-
-        prob = analysis_data.get('acceptance_probability')
-        
+        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'][:5]
+        acceptance_prob = analysis_data.get('acceptance_probability', None)
+        acceptance_line = ""
+        if acceptance_prob is not None:
+            acceptance_line = f"\n*Hiring Potential Score: {acceptance_prob:.2f}%*\n"
+            if acceptance_prob >= 80: acceptance_line += "Assessment: Exceptional candidate, strongly recommended for advancement."
+            elif acceptance_prob >= 50: acceptance_line += "Assessment: Promising candidate with moderate strengths; consider for further evaluation."
+            else: acceptance_line += "Assessment: Limited alignment with role expectations; significant development needed."
         prompt = f"""
-        *Persona:* You are a Senior HR Partner writing a candidate evaluation memo for the hiring manager.
-        *Task:* Write a professional, objective, and concise evaluation based on the data below.
-        *Tone:* Analytical and formal.
-
-        *CANDIDATE EVALUATION MEMORANDUM*
-        *CONFIDENTIAL*
-
-        *Candidate ID:* {analysis_data.get('user_id', 'N/A')}
-        *Analysis Date:* {time.strftime('%Y-%m-%d')}
-        *Estimated Acceptance Probability:* {prob:.2f}%
-
-        *1. Overall Recommendation:*
-        Provide a clear, one-sentence recommendation (e.g., "Highly recommend proceeding to the final round," or "Recommend with reservations due to...").
-
-        *2. Key Competency Assessment (Content & Skills):*
-        - Summarize the candidate's key strengths and areas for development based on the content analysis.
-        - *Data for Content Analysis:* {content_interp}
-
-        *3. Communication Style (Voice & Speech Analysis):*
-        - Evaluate the candidate's communication style (confidence, clarity, nervousness).
-        - *Data for Voice Analysis:* {voice_interp}
-        
-        *4. Actionable Next Steps:*
-        - Suggest specific questions or topics for the next interviewer to focus on.
+        You are an expert HR consultant, EvalBot, tasked with producing a professional, concise, and actionable interview analysis report. Structure the report with clear headings, subheadings, and bullet points (use '- ' for bullets). Adopt a formal, HR-professional tone, focusing on candidate evaluation, fit for role, and development insights.
+        {acceptance_line}
+        *1. Executive Summary*
+        - Provide a concise overview of the interview, highlighting key metrics and overall candidate performance.
+        - Interview duration: {analysis_data['text_analysis']['total_duration']:.2f} seconds
+        - Total speaker turns: {analysis_data['text_analysis']['speaker_turns']}
+        - Participants: {', '.join(analysis_data['speakers'])}
+        *2. Communication and Vocal Analysis*
+        - Evaluate the candidate's vocal delivery, including speaking rate, fluency, and confidence indicators.
+        - Provide HR-relevant insights into how these metrics impact perceived professionalism and role suitability.
+        {voice_interpretation}
+        *3. Content Analysis and Competency Assessment*
+        - Analyze key themes in the candidate's responses to assess alignment with job competencies (e.g., problem-solving, communication, leadership).
+        - Identify strengths and areas for improvement, supported by specific examples.
+        - Sample responses for context:
+        {chr(10).join(interviewee_responses)}
+        *4. Fit and Potential Evaluation*
+        - Assess the candidate's overall fit for a typical professional role based on communication, content, and vocal dynamics.
+        - Consider cultural fit, adaptability, and readiness for the role.
+        *5. Actionable HR Recommendations*
+        - Provide specific, prioritized recommendations for the candidate’s development.
+        - Focus areas: Effective Communication, Content Clarity and Depth, Professional Presence.
+        - Suggest next steps for hiring managers (e.g., advance to next round, additional assessments, training focus).
         """
         response = gemini_model.generate_content(prompt)
         return response.text
@@ -334,63 +475,156 @@ def generate_report(analysis_data: Dict) -> str:
         logger.error(f"Report generation failed: {str(e)}")
         return f"Error generating report: {str(e)}"
 
-# --- NEW: Polished PDF Creation ---
-def parse_gemini_report(text: str) -> list:
-    parsed_elements = []
-    patterns = {
-        'h3': r'^\s*\\\d\.\d\s+(.?)\\*:',
-        'bullet': r'^\s*[-•]\s(.*)',
-        'bold': r'^\s*\\(.?)\\*'
-    }
-    for line in text.split('\n'):
-        line = line.strip()
-        if not line: continue
-        match_h3 = re.match(patterns['h3'], line)
-        if match_h3:
-            parsed_elements.append({'type': 'h3', 'content': match_h3.group(1)})
-            continue
-        match_bold = re.match(patterns['bold'], line)
-        if match_bold:
-            if not re.match(r'^\d\.', match_bold.group(1)):
-                 parsed_elements.append({'type': 'h3', 'content': match_bold.group(1)})
-                 continue
-        match_bullet = re.match(patterns['bullet'], line)
-        if match_bullet:
-            parsed_elements.append({'type': 'bullet', 'content': match_bullet.group(1)})
-            continue
-        parsed_elements.append({'type': 'body', 'content': line})
-    return parsed_elements
-
 def create_pdf_report(analysis_data: Dict, output_path: str, gemini_report_text: str):
     try:
-        doc = SimpleDocTemplate(output_path, pagesize=letter, rightMargin=0.75*inch, leftMargin=0.75*inch, topMargin=1.2*inch, bottomMargin=1*inch)
+        doc = SimpleDocTemplate(output_path, pagesize=letter,
+                                rightMargin=0.75*inch, leftMargin=0.75*inch,
+                                topMargin=1*inch, bottomMargin=1*inch)
         styles = getSampleStyleSheet()
-        h1 = ParagraphStyle(name='Heading1', fontSize=18, leading=22, spaceAfter=12, alignment=1, textColor=colors.HexColor('#00205B'), fontName='Helvetica-Bold')
-        h2 = ParagraphStyle(name='Heading2', fontSize=14, leading=18, spaceBefore=18, spaceAfter=10, textColor=colors.HexColor('#003366'), fontName='Helvetica-Bold')
-        h3 = ParagraphStyle(name='Heading3', parent=h2, fontSize=11, spaceBefore=10, spaceAfter=4, textColor=colors.HexColor('#2E8B57'), fontName='Helvetica-Bold')
-        body_text = ParagraphStyle(name='BodyText', parent=styles['Normal'], fontSize=10, leading=14, spaceAfter=6, fontName='Helvetica')
-        bullet_style = ParagraphStyle(name='Bullet', parent=body_text, leftIndent=20, bulletIndent=10, spaceAfter=4)
+        h1 = ParagraphStyle(name='Heading1', fontSize=22, leading=26, spaceAfter=20, alignment=1, textColor=colors.HexColor('#1A3C5E'))
+        h2 = ParagraphStyle(name='Heading2', fontSize=14, leading=18, spaceBefore=14, spaceAfter=8, textColor=colors.HexColor('#2E5A87'))
+        body_text = ParagraphStyle(name='BodyText', parent=styles['Normal'], fontSize=10, leading=14, spaceAfter=8, fontName='Helvetica')
+        bullet_style = ParagraphStyle(name='Bullet', parent=body_text, leftIndent=20, bulletIndent=10, fontName='Helvetica')
+        
         story = []
+
         def header_footer(canvas, doc):
             canvas.saveState()
             canvas.setFont('Helvetica', 9)
             canvas.setFillColor(colors.grey)
-            canvas.drawString(doc.leftMargin, 0.5 * inch, f"Page {doc.page} | EvalBot Confidential Report")
-            canvas.setStrokeColor(colors.HexColor('#003366'))
-            canvas.setLineWidth(0.5)
-            canvas.line(doc.leftMargin, doc.height + 0.8*inch, doc.width + doc.leftMargin, doc.height + 0.8*inch)
+            canvas.drawString(doc.leftMargin, 0.5 * inch, f"Page {doc.page} | EvalBot HR Interview Report | Confidential")
+            canvas.setStrokeColor(colors.HexColor('#2E5A87'))
+            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.setFillColor(colors.HexColor('#003366'))
-            canvas.drawString(doc.leftMargin, doc.height + 0.9*inch, "Interview Performance Analysis")
+            canvas.drawString(doc.leftMargin, doc.height + 0.9*inch, "Candidate Interview Analysis Report")
             canvas.restoreState()
 
-        # Build the story from the parsed Gemini report
-        parsed_report = parse_gemini_report(gemini_report_text)
-        for element in parsed_report:
-            if element['type'] == 'h2': story.append(Paragraph(element['content'], h2))
-            elif element['type'] == 'h3': story.append(Paragraph(element['content'], h3))
-            elif element['type'] == 'bullet': story.append(Paragraph(f"• {element['content']}", bullet_style))
-            else: story.append(Paragraph(element['content'], body_text))
+        # Title Page
+        story.append(Paragraph("Candidate Interview Analysis Report", h1))
+        story.append(Paragraph(f"Generated on: {time.strftime('%B %d, %Y')}", ParagraphStyle(name='Date', alignment=1, fontSize=10, textColor=colors.grey)))
+        story.append(Spacer(1, 0.5 * inch))
+        acceptance_prob = analysis_data.get('acceptance_probability')
+        if acceptance_prob is not None:
+            story.append(Paragraph("Hiring Potential Snapshot", h2))
+            prob_color = colors.HexColor('#2E7D32') if acceptance_prob >= 70 else (colors.HexColor('#F57C00') if acceptance_prob >= 40 else colors.HexColor('#D32F2F'))
+            story.append(Paragraph(f"Hiring Potential Score: <font size=16 color='{prob_color.hexval()}'><b>{acceptance_prob:.2f}%</b></font>",
+                                 ParagraphStyle(name='Prob', fontSize=12, spaceAfter=12, alignment=1)))
+            if acceptance_prob >= 80:
+                story.append(Paragraph("<b>HR Assessment:</b> Exceptional candidate, strongly recommended for advancement to the next stage.", body_text))
+            elif acceptance_prob >= 50:
+                story.append(Paragraph("<b>HR Assessment:</b> Promising candidate with moderate strengths; consider for further evaluation.", body_text))
+            else:
+                story.append(Paragraph("<b>HR Assessment:</b> Limited alignment with role expectations; significant development needed.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+        story.append(Paragraph("Prepared by: EvalBot - AI-Powered HR Interview Analysis System", body_text))
+        story.append(PageBreak())
+
+        # Detailed Analysis
+        story.append(Paragraph("Detailed Candidate Evaluation", h1))
+        
+        story.append(Paragraph("1. Communication and Vocal Profile", 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", 'Optimal: 2.0-3.0 wps; impacts clarity and confidence'],
+                ['Filler Word Usage', f"{voice_analysis.get('filler_ratio', 0) * 100:.1f}%", 'High usage may reduce perceived professionalism'],
+                ['Anxiety Indicator', voice_analysis.get('interpretation', {}).get('anxiety_level', 'N/A'), f"Score: {voice_analysis.get('composite_scores', {}).get('anxiety', 0):.3f}; reflects pressure response"],
+                ['Confidence Indicator', voice_analysis.get('interpretation', {}).get('confidence_level', 'N/A'), f"Score: {voice_analysis.get('composite_scores', {}).get('confidence', 0):.3f}; indicates vocal authority"],
+                ['Fluency Assessment', voice_analysis.get('interpretation', {}).get('fluency_level', 'N/A'), 'Affects engagement and message delivery']
+            ]
+            table = Table(table_data, colWidths=[1.8*inch, 1.2*inch, 3.5*inch])
+            table.setStyle(TableStyle([
+                ('BACKGROUND', (0,0), (-1,0), colors.HexColor('#2E5A87')),
+                ('TEXTCOLOR', (0,0), (-1,0), colors.whitesmoke),
+                ('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), 12),
+                ('TOPPADDING', (0, 0), (-1, 0), 12),
+                ('BACKGROUND', (0, 1), (-1, -1), colors.HexColor('#F5F7FA')),
+                ('GRID', (0,0), (-1,-1), 1, colors.HexColor('#DDE4EB'))
+            ]))
+            story.append(table)
+            story.append(Spacer(1, 0.25 * 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=4.5*inch, height=2.8*inch)
+            img.hAlign = 'CENTER'
+            story.append(img)
+        else:
+            story.append(Paragraph("Voice analysis unavailable due to processing limitations.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+
+        # Parse Gemini Report
+        sections = {}
+        section_titles = ["Executive Summary", "Communication and Vocal Analysis", 
+                         "Content Analysis and Competency Assessment", 
+                         "Fit and Potential Evaluation", "Actionable HR Recommendations"]
+        for title in section_titles:
+            sections[title] = []
+        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 section_titles:
+                if title.lower() in part.lower():
+                    current_section = title
+                    is_heading = True
+                    break
+            if not is_heading and current_section:
+                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 not available from analysis.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+
+        # Content and Competency
+        story.append(Paragraph("3. Content and Competency Assessment", h2))
+        if sections['Content Analysis and Competency Assessment']:
+            for line in sections['Content Analysis and Competency Assessment']:
+                if line.startswith(('-', '•', '*')):
+                    story.append(Paragraph(line.lstrip('-•* ').strip(), bullet_style))
+                else:
+                    story.append(Paragraph(line, body_text))
+        else:
+            story.append(Paragraph("Content and competency analysis not provided.", body_text))
+        story.append(PageBreak())
+
+        # Fit and Potential
+        story.append(Paragraph("4. Fit and Potential Evaluation", h2))
+        if sections['Fit and Potential Evaluation']:
+            for line in sections['Fit and Potential Evaluation']:
+                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 evaluation not available.", body_text))
+        story.append(Spacer(1, 0.3 * inch))
+
+        # HR Recommendations
+        story.append(Paragraph("5. Actionable HR Recommendations", h2))
+        if sections['Actionable HR Recommendations']:
+            for line in sections['Actionable HR Recommendations']:
+                if line.startswith(('-', '•', '*')):
+                    story.append(Paragraph(line.lstrip('-•* ').strip(), bullet_style))
+                else:
+                    story.append(Paragraph(line, body_text))
+        else:
+            story.append(Paragraph("HR recommendations not provided.", body_text))
 
         doc.build(story, onFirstPage=header_footer, onLaterPages=header_footer)
         return True
@@ -398,81 +632,64 @@ def create_pdf_report(analysis_data: Dict, output_path: str, gemini_report_text:
         logger.error(f"Enhanced PDF creation failed: {str(e)}", exc_info=True)
         return False
 
+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
 
-# --- MAIN ORCHESTRATOR FUNCTION ---
 def process_interview(audio_path_or_url: str):
-    local_audio_path, wav_file, is_downloaded = None, None, False
+    local_audio_path = None
+    wav_file = None
+    is_downloaded = False
     try:
-        user_id_from_task = "unknown_user"
-        try:
-            from celery_worker import celery_app
-            if celery_app.current_task:
-                user_id_from_task = celery_app.current_task.request.kwargs.get('item_data', {}).get('user_id', 'unknown_user')
-        except (ImportError, AttributeError):
-            pass # Celery might not be in the context if run locally
-
         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)
         transcript = transcribe(wav_file)
-        
-        for u in transcript['utterances']:
-            u['prosodic_features'] = extract_prosodic_features(wav_file, u['start'], u['end'])
-            
+        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)
-        
-        # NOTE: Using alternating role classification as decided.
-        for i, u in enumerate(utterances_with_speakers):
-            u['role'] = 'Interviewer' if i % 2 == 0 else 'Interviewee'
-        classified_utterances = utterances_with_speakers
-        
+        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(utterances_with_speakers)
+        classified_utterances = classify_roles(utterances_with_speakers, clf, vectorizer, scaler)
         voice_analysis = analyze_interviewee_voice(wav_file, classified_utterances)
-        content_analysis = analyze_text_content(classified_utterances)
-        
         analysis_data = {
-            'user_id': user_id_from_task,
             'transcript': classified_utterances,
             'speakers': list(set(u['speaker'] for u in classified_utterances)),
             'voice_analysis': voice_analysis,
-            'advanced_content_analysis': content_analysis,
             'text_analysis': {
                 '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)
-        
         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)
-        
+        create_pdf_report(analysis_data, pdf_path, gemini_report_text=gemini_report_text)
         with open(json_path, 'w') as f:
-            json.dump(convert_to_serializable(analysis_data), f, indent=2)
-            
+            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,
-            'pdf_filename': os.path.basename(pdf_path),
-            'json_filename': os.path.basename(json_path)
-        }
-
+        return {'pdf_path': pdf_path, 'json_path': json_path}
     except Exception as e:
         logger.error(f"Processing failed for {audio_path_or_url}: {str(e)}", exc_info=True)
         raise
-
     finally:
-        if wav_file and os.path.exists(wav_file): os.remove(wav_file)
+        if wav_file and os.path.exists(wav_file):
+            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}")