🎙️ Talklytics

Real-Time Conversational Interest Detection System

Talklytics analyses live sales calls and outputs a Convincingness Score (1–5) that reflects how interested a customer sounds — using both what they say and how they say it.

---

Architecture

                     Microphone Input
                            ↓
                AudioCapture (sounddevice)
                            ↓
         AudioPreprocessor — normalise + silence filter
                            ↓
               ┌────────────┴────────────┐
               │                         │
       AcousticExtractor         SpeechRecognizer (Faster-Whisper)
       (librosa: MFCC,                   ↓
        pitch, energy)           LinguisticAnalyzer (DistilBERT)
               │                 (sentiment, keywords)
               └────────────┬────────────┘
                            ↓
               FusionModel (Random Forest)
                            ↓
             EngagementTracker → Streamlit Dashboard

---

Quick Start

# 1. Create a virtual environment
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate

# 2. Install dependencies
pip install -r requirements.txt

# 3. Launch the dashboard
streamlit run dashboard/app.py

# — or — run headless in the terminal
python main.py

The dashboard is served at http://localhost:8501

---

Training the Fusion Model

# Using RAVDESS
python training/train_model.py --ravdess data/ravdess

# Using a custom CSV (columns: audio_path, transcript, score)
python training/train_model.py --csv data/labeled_calls.csv

Without a trained model the system uses a built-in heuristic fallback.

---

Project Structure

Talklytics-update/
├── config.py                   ← All tunable parameters
├── main.py                     ← Headless entry point
├── requirements.txt
├── modules/
│   ├── audio_capture.py        ← SoundDevice microphone streaming
│   ├── audio_preprocessor.py  ← Normalisation & silence filter
│   ├── acoustic_extractor.py  ← Librosa: MFCC, pitch, energy
│   ├── speech_recognizer.py   ← Faster-Whisper ASR
│   ├── linguistic_analyzer.py ← DistilBERT sentiment + keywords
│   ├── fusion_model.py        ← Random Forest classifier
│   ├── engagement_tracker.py  ← Timeline of engagement records
│   └── pipeline.py            ← Orchestrator (background thread)
├── dashboard/
│   └── app.py                 ← Streamlit real-time dashboard
├── training/
│   ├── data_loader.py         ← RAVDESS / CSV loaders
│   └── train_model.py         ← CLI training script
├── models/                    ← Saved model artifacts (.pkl)
└── data/                      ← Place datasets here

---

Score Scale

Score	Label
1	Disengaged
2	Low Interest
3	Neutral
4	Interested
5	Highly Interested

---

Technology Stack

Layer	Technology
Language	Python 3.9+
Audio Capture	sounddevice
Audio Processing	Librosa
Speech Recognition	Faster-Whisper
NLP	HuggingFace Transformers
Machine Learning	Scikit-Learn
Data Processing	NumPy, Pandas
Visualisation	Matplotlib
Dashboard	Streamlit

---

Hardware Requirements

Component	Minimum
CPU	Intel i5 / AMD equiv.
RAM	8 GB
Storage	10 GB free
GPU	Optional

No GPU required — the system runs entirely on CPU.