Resonance

Transforms speech into real-time emotional insights. We transcribe voice, analyze tone and sentiment, and visualize emotional trends to help teams communicate better and detect stress early.

Comment

Our Story:

Multi-speaker conversations often contain subtle emotional dynamics and recurring patterns that are difficult to identify from raw transcripts. Traditional transcripts provide text but fail to reveal:

  • Tension points and conflict areas
  • Repeated arguments or looping topics
  • Conversational imbalance between speakers

Thus, we built Resonance, a tool that transforms raw audio into structured emotional intelligence, making it easy to understand not just what was said, but what it means emotionally.

1. Audio Upload & Processing

  • Supports multiple formats: pre-recorded or live recording
  • Real-time upload progress tracking
  • Job-based processing system with status updates

2. Speech Recognition & Diarization

  • Automatic speaker diarization (identifies who said what)
  • Utterance-level segmentation for granular insights
  • High-accuracy transcription with punctuation and smart formatting

3. Sentiment Analysis

  • Per-utterance sentiment scoring (normalized between [-1, 1]).
  • Real-time sentiment visualization across the conversation.
  • Conflict heatmap showing negative sentiment intensity.
  • Sentiment "calculus" charts for trend analysis.

4. Analytics Dashboard

  • Transcript View: Speaker-colored transcript with clickable timestamps.
  • Sentiment Visualization: Interactive charts showing sentiment over time.
  • Conflict Heatmap: Visual representation of conflict intensity.
  • Insight Cards: Clickable insights that jump to specific timestamps.
  • Speaker Analysis: Detailed statistics per speaker (talk time, turns, average sentiment).
  • Topics & Intents: Extracted conversation topics and speaker intents.
  • Audio Playback: Synchronized audio playback with transcript highlighting.
  • CSV Export: Download analysis data for further processing.

Resonance is a conversation intelligence platform that analyzes multi-speaker audio recordings to reveal emotional dynamics, sentiment patterns, and conversational insights, allowing professionals and non-professionals alike to focus on what matters most- connecting with their clients personally.

How We Excel:

  • Diverse input formats: Upload any multi-speaker conversation and receive structured emotional and conversational analysis.
  • Real-time sentiment tracking: Visualize emotional trajectories over time with per-utterance sentiment analysis.
  • Composite detection: Automatically identify tension points, conflict spikes, and emotional shifts in conversations.
  • Speaker intelligence: Get speaker-diarized transcripts with talk time, turn count, and sentiment per speaker.
  • Agentic NL analysis: Leverage Deepgram's Nova-3 & Gemini 3 Flash for advanced speech recognition & conversational understanding.
  • Privacy-focused: Audio files are processed in our client-side cache, where user privacy is our priority.

User Groups:

  • For Management: Understand team dynamics, identify conflict points, track sentiment throughout discussions
  • For Research: Extract structured insights from qualitative interviews, analyze sentiment and topics
  • For Care Providers: Gain self-awareness by analyzing personal conversations and emotional patterns
  • For You: Where you deserve the value from every single meaningful interaction.

Tech Team:

  • Nakul Soneji (NakulSoneji, UC Berkeley)
  • Eashan Mahajan (TempsL8, UC Berkeley)
  • Ethan Sie (ethannsie, UC Berkeley)
  • Kevin Geng (KevinGeng07, UC Berkeley)

Development Journey:

The project started with a clear Product Requirements Document (PRD) and Technical Design Document (TDD) outlining the vision. The implementation followed a batch processing architecture where:

  1. Users upload audio files
  2. Backend processes files through Deepgram API
  3. Results are analyzed for insights
  4. Dashboard displays comprehensive visualizations

The system was designed with scalability in mind, using job-based processing that can handle multiple concurrent uploads. The frontend was built with modern UX patterns, ensuring a smooth user experience with loading states, error handling, and real-time backend updates.

The Future:

  • Comparative session & real-time streaming analytics.
  • Speaker correction interface w/manual speaker relabeling.
  • Long-term conversation history storage & enhanced confidence displays.

How We're Built:

  • Frontend: React 18 with TypeScript, using modern UI components (Radix UI, shadcn/ui)
  • Backend: Express.js (Node.js) with TypeScript
  • Database: PostgreSQL with Drizzle ORM for job storage and user management
  • APIs:
    • Deepgram API for speech-to-text, diarization, and sentiment analysis
    • Google Gemini API for AI-powered summaries and chat
  • State Management: TanStack Query (React Query) for server state
  • Styling: Tailwind CSS with custom theme support (light/dark mode)
  • Build Tools: Vite for fast development and optimized production builds
  • Authentication: Passport.js with local strategy, session-based auth

The Frontend:

  • React - Modern UI library for building interactive interfaces
  • TypeScript - Type-safe JavaScript for better code quality
  • Vite - Next-generation frontend build tool for fast development
  • Wouter - Lightweight routing library
  • TanStack Query (React Query) - Powerful data synchronization for React
  • Framer Motion - Production-ready motion library for animations
  • Recharts - Composable charting library for data visualization
  • Tailwind CSS - Utility-first CSS framework
  • Radix UI - Unstyled, accessible component primitives:
    • Accordion, Alert Dialog, Avatar, Checkbox, Dialog, Dropdown Menu
    • Popover, Progress, Radio Group, Select, Slider, Switch, Tabs
    • Toast, Tooltip, and many more UI components
  • Lucide React - Beautiful icon library
  • Zod - TypeScript-first schema validation
  • Next Themes - Theme switching (light/dark mode)

The Backend:

  • Node.js - JavaScript runtime environment
  • Express.js - Fast, unopinionated web framework
  • Passport.js - Authentication middleware
  • TypeScript - Type-safe backend development
  • PostgreSQL - Relational database for job storage and user management
  • Drizzle ORM - TypeScript ORM for database operations
  • Multer - Middleware for handling multipart/form-data (file uploads)
  • WebSocket - Real-time communication support

The APIs & Services:

  • Deepgram Nova-3 - Speech-to-text, speaker diarization, and sentiment analysis
  • Google Gemini 3 Flash - AI summaries, interactive chat, context-aware responses

Misc. & Deployment:

  • ESBuild - Fast JavaScript bundler
  • Drizzle Kit - Database migrations and introspection
  • PostCSS - CSS processing
  • Autoprefixer - CSS vendor prefixing

Challenges:

Pivoting product development from simple sentiment understanding to a pre-production multi-faceted conversational intelligence platform, all in under 24 hours. This pivot challenged us with a mountain of new technology stacks to adapt to in order to create a robust representation of our envisioned product.

Accomplishments:

Communicating with clients, from corporate directors, to therapists, to middle Americans, we scheduled and conducted 10-minute blitz interviews to understand their pain points in conversational intelligence. Through it, we refined our definition of technology that fixes, rather than muddles, their interactions with others. Through this process, we also realized how we could drive confident, yet cognizant, interactions.

Built With

Share this project:

Updates