Sonix

Audio-Based Conversation Analysis

🎧 Sonix — Audio-Based Conversation Analysis

Sonix is a Python library designed to extract rich analytical signals directly from audio files — without relying on transcripts or text analysis. It focuses purely on acoustic and prosodic features to help researchers, developers, and data scientists understand conversational dynamics, emotional tone, and speaking patterns.

---

🚀 Features

Sonix provides end-to-end analysis of raw audio conversations, including:

Category	Description	Example Metrics
🎚️ Basic Audio Stats	Extracts simple sound metrics useful for quality and consistency checks.	RMS Energy, Duration, Silence Ratio
🗣️ Voice Activity Detection (VAD)	Identifies speaking vs. silence segments.	Speech Segments, Turn Counts
🎵 Pitch & Prosody	Analyzes intonation and variation in tone.	Average Pitch, Pitch Variance
💬 Speech Tempo	Measures speaking rate and rhythm.	Words per Minute (approx), Speech Rate
🔊 Energy Dynamics	Examines loudness variation to detect emphasis or excitement.	Mean Energy, Energy Variability
😠 Emotion & Tone Estimation	Classifies emotional states using pretrained acoustic models.	Calm, Happy, Angry, Sad
⏱️ Overlap & Turn-Taking	Detects interruptions and conversational overlap.	Speaker Overlap %, Turn Durations
🌈 Spectral Features	Extracts frequency-domain data for ML and acoustic analysis.	MFCCs, Spectral Centroid, Roll-off
🎯 Audio Quality	Evaluates clarity and background noise.	SNR (Signal-to-Noise Ratio), Distortion
🧠 Derived Conversation Insights	Combines features to infer interaction quality.	Engagement Index, Talk/Listen Ratio

---

🧩 Example Use Cases

• Conversation analytics for call centers or AI voice agents
• Measuring emotional tone or stress levels in speech
• Detecting dominance or interruptions in meetings
• Generating audio-based KPIs for human–AI interactions
• Building real-time feedback tools for voice communication training

---

🛠️ Installation

pip install Sonix

---

🧪 Quick Start

from Sonix import AudioAnalyzer

# Load and analyze an audio file
analyzer = AudioAnalyzer("conversation.wav")

# Run full analysis
report = analyzer.analyze_all()

# Print summary
print(report.summary())

# Access individual feature groups
print(report.pitch.mean)
print(report.energy.variance)
print(report.emotion.probabilities)

---

📊 Example Output

{
  "duration_sec": 180.4,
  "speech_segments": 56,
  "average_pitch_hz": 201.3,
  "pitch_variance": 32.8,
  "mean_energy": -20.5,
  "energy_variability": 0.17,
  "speech_rate_wpm": 142,
  "emotion": {
    "calm": 0.52,
    "happy": 0.28,
    "angry": 0.12,
    "sad": 0.08
  },
  "overlap_ratio": 0.07,
  "engagement_index": 0.81
}

---

⚙️ Architecture Overview

Sonix/
│
├── core/
│   ├── audio_loader.py         # Handles input normalization, channel merging
│   ├── feature_extractor.py    # Extracts MFCCs, pitch, energy, etc.
│   ├── vad_detector.py         # Voice activity segmentation
│   ├── prosody_analyzer.py     # Pitch, tone, tempo analysis
│   ├── emotion_estimator.py    # Acoustic emotion classification
│   ├── quality_metrics.py      # Noise and clarity estimation
│   └── report_builder.py       # Combines results into structured JSON
│
├── models/
│   ├── emotion_model.onnx
│   └── vad_model.onnx
│
└── utils/
    ├── visualization.py        # Waveform and spectrum plots
    └── audio_helpers.py

---

🧬 Technical Dependencies

Library	Purpose
librosa	Audio feature extraction
pyAudioAnalysis	Energy and tempo metrics
webrtcvad	Voice activity detection
torch / onnxruntime	Emotion classification models
numpy / scipy	Signal processing
matplotlib	Visualization (optional)

---

🧠 Example Analysis Pipeline

from Sonix.pipeline import AudioPipeline

pipeline = AudioPipeline([
  "vad",
  "pitch",
  "energy",
  "emotion",
  "turn_taking"
])

results = pipeline.run("meeting.wav")
results.plot_waveform()

---

📈 Metrics Summary

Metric	Description
duration_sec	Total length of the audio file
speech_segments	Number of detected speaking parts
speech_rate_wpm	Approximate speech tempo
average_pitch_hz	Average fundamental frequency
energy_variability	Standard deviation of loudness
overlap_ratio	% of overlapping speech between speakers
emotion	Probabilities of acoustic emotion categories
engagement_index	Combined measure of voice energy, tempo, and tone consistency

---

🧰 CLI Usage

Sonix analyze conversation.wav --plot

Output includes JSON summary + waveform visualization.

---

🧑‍💻 Example Integration

Integrate with your AI or analytics platform:

from Sonix import AudioAnalyzer

analyzer = AudioAnalyzer("agent_call.wav")
signals = analyzer.get_signals()
agent_metrics = {
  "engagement": signals["engagement_index"],
  "emotion": signals["emotion"]["happy"],
  "speech_rate": signals["speech_rate_wpm"]
}

---

🔍 Future Roadmap

• Real-time streaming analysis
• Speaker diarization and identification
• Gender and age acoustic profiling
• Conversation quality score model
• REST API integration

---

🧑‍🔬 Author

Developed by Dima Statz & Contributors
📫 Contributions welcome via pull requests and GitHub issues.

---

📄 License

MIT License © 2025 Sonix Team