#!/usr/bin/env python3 import sys import json import traceback import re def count_syllables(word): word = word.lower() count = 0 vowels = "aeiouy" if len(word) == 0: return 0 if word[0] in vowels: count += 1 for index in range(1, len(word)): if word[index] in vowels and word[index - 1] not in vowels: count += 1 if word.endswith("e"): count -= 1 if count == 0: count = 1 return count def compute_flesch_kincaid(text): # Basic word tokenization words = re.findall(r'\b\w+\b', text) if not words: return 0.0 # Sentence splitting (approximate) sentences = re.split(r'[.!?]+', text) sentences = [s.strip() for s in sentences if s.strip()] if not sentences: sentences = [text] num_words = len(words) num_sentences = len(sentences) num_syllables = sum(count_syllables(w) for w in words) asl = num_words / num_sentences asw = num_syllables / num_words grade = 0.39 * asl + 11.8 * asw - 15.59 return round(max(0.0, min(18.0, grade)), 1) def main(): try: # Read JSON parameters from stdin input_data = json.load(sys.stdin) transcript = input_data.get('transcript', '') # 1. Load NLTK and initialize VADER import nltk from nltk.sentiment.vader import SentimentIntensityAnalyzer # Initialize VADER sia = SentimentIntensityAnalyzer() # Split into sentences using nltk sentences = nltk.sent_tokenize(transcript) sentiment_arc = [] for idx, sentence in enumerate(sentences): sentence = sentence.strip() if not sentence: continue # Get VADER compound polarity score (-1.0 to +1.0) scores = sia.polarity_scores(sentence) compound = scores['compound'] # Context-aware mapping using standard thresholds if compound >= 0.05: label = 'POSITIVE' elif compound <= -0.05: label = 'NEGATIVE' else: label = 'NEUTRAL' # Normalize compound score from [-1.0, 1.0] to [0.0, 1.0] normalized_score = float((compound + 1.0) / 2.0) sentiment_arc.append({ 'label': label, 'score': normalized_score, 'sentence': sentence, 'sentence_index': idx }) if not sentiment_arc: sentiment_arc.append({ 'label': 'NEUTRAL', 'score': 0.5, 'sentence': 'No speech recorded.', 'sentence_index': 0 }) # 2. Semantic concepts similarity using Sentence-Transformers from sentence_transformers import SentenceTransformer, util # Use a lightweight fast embedding model model = SentenceTransformer('all-MiniLM-L6-v2') concepts = [ { 'concept': 'Value Proposition', 'target_statements': [ 'we offer automations to improve and speed up workflows', 'solves delays in shipping and coordinates partners', 'increases efficiency and value for companies' ] }, { 'concept': 'Pricing & Budget', 'target_statements': [ 'pricing plans subscription budgets and setup costs', 'how much it costs affordable pricing packages and investment return', 'budget options and financial details of onboarding' ] }, { 'concept': 'Implementation', 'target_statements': [ 'onboarding setup timeframes system integrations', 'launching the software setting up database connections', 'how to implement get started training the team' ] }, { 'concept': 'Closing Statement', 'target_statements': [ 'schedule next steps and book a follow up demo meeting', 'sign contract get started next week and closing the deal', 'call to action setting up a calendar invitation' ] } ] sales_concepts = [] if transcript.strip() and sentences: # Embed entire transcript sentences for matching sentence_embeddings = model.encode(sentences, convert_to_tensor=True) for c in concepts: # Embed target concept statements target_embeddings = model.encode(c['target_statements'], convert_to_tensor=True) # Compute similarity matrix between all sentences and target statements sim_matrix = util.cos_sim(sentence_embeddings, target_embeddings) # Find the best matching sentence indices and scores max_sim = float(sim_matrix.max()) detected = max_sim >= 0.40 evidence_sentence = "" if detected: # Get index of sentence with highest similarity import numpy as np flat_idx = int(sim_matrix.cpu().numpy().argmax()) sent_idx = flat_idx // len(c['target_statements']) if sent_idx < len(sentences): evidence_sentence = sentences[sent_idx] sales_concepts.append({ 'concept': c['concept'], 'detected': detected, 'confidence': max_sim, 'evidence_sentence': evidence_sentence }) else: for c in concepts: sales_concepts.append({ 'concept': c['concept'], 'detected': False, 'confidence': 0.0, 'evidence_sentence': "" }) # 3. Readability & Power Words flesch_grade = compute_flesch_kincaid(transcript) # Match power words power_words_list = [ 'value', 'revenue', 'growth', 'solution', 'results', 'guarantee', 'discover', 'proven', 'save', 'efficient', 'optimize', 'speed', 'increase', 'improve', 'maximize', 'seamless', 'instant' ] lower_transcript = transcript.lower() power_word_hits = [] for pw in power_words_list: if re.search(r'\b' + re.escape(pw) + r'\b', lower_transcript): power_word_hits.append(pw.capitalize()) # Match top keywords (highest frequency nouns/adjectives) from nltk.corpus import stopwords try: stop_words = set(stopwords.words('english')) except Exception: # Download stopwords if not present nltk.download('stopwords') stop_words = set(stopwords.words('english')) words_filtered = [w for w in re.findall(r'\b\w+\b', lower_transcript) if w not in stop_words and len(w) > 2] # Count frequencies freq = {} for w in words_filtered: freq[w] = freq.get(w, 0) + 1 top_keywords = sorted(freq, key=freq.get, reverse=True)[:5] top_keywords = [w.capitalize() for w in top_keywords] # Output results result = { 'status': 'success', 'sentiment_arc': sentiment_arc, 'sales_concepts': sales_concepts, 'flesch_kincaid_grade': flesch_grade, 'power_word_hits': power_word_hits, 'top_keywords': top_keywords } print(json.dumps(result)) except Exception as e: error_details = { 'status': 'error', 'error': str(e), 'traceback': traceback.format_exc() } print(json.dumps(error_details), file=sys.stderr) sys.exit(1) if __name__ == '__main__': main()