🎙️ Oratio — Voice Note Intelligence & Spoken Delivery Coach

Turn informal voice memos into high-precision, timestamped speaking telemetry and personalized AI coaching.

💡 Inspiration

In today's remote and asynchronous work culture, the highest-stakes communication rarely happens in 30-minute keynote presentations with rehearsed slides. It happens in 90-second voice notes to a co-founder, impromptu sprint standup updates, and voice messages to clients or investors.

Yet, existing speech-coaching tools are archaic: they are heavy desktop overlays built exclusively for scheduled Zoom meetings or scripted speeches. They provide vague, retroactive scores like "Clarity: 7/10", but they cannot tell you:

When your cadence spiked under anxiety,
Where conversational filler crutches compounded, or
How dead-air stalls interrupted your logical flow.

We built Oratio to bridge this gap: a privacy-first, web-native voice intelligence platform that analyzes spontaneous spoken delivery through acoustic signal telemetry and grounds actionable feedback directly onto audio timestamps.

🛠️ What It Does

Oratio operates as a personal voice intelligence studio and delivery coach:

Dual Intake Studio:
- Record Live: Ingest audio with real-time browser microphone capture, live waveform visualization, and instantaneous client-side speech recognition.
- Audio File Upload: Upload uncompressed or lossy voice recordings (.mp3, .wav, .m4a) from WhatsApp, Telegram, or phone voice memos.
Acoustic Telemetry & Temporal Grounding:
- Computes delivery metrics against a personalized baseline: $$\text{WPM} = \frac{\text{Total Word Count}}{\text{Duration in Minutes}} = \frac{W}{\Delta t / 60}$$ $$\text{Filler Density} = \left( \frac{\text{Count of Fillers}}{\text{Total Word Count}} \right) \times 100\%$$ $$\text{Silence Ratio} = \left( \frac{\sum \text{Pause Durations}}{\text{Total Duration}} \right) \times 100\%$$
- Flags temporal delivery flaws (pace spikes $> 175\text{ WPM}$, filler clusters, dead-air pauses $> 1.5\text{s}$) with interactive timestamp anchors on the audio scrub bar.
Multi-Factor Semantic Rubric:
- Evaluates delivery, clarity, structure, content, fluency, and engagement on a weighted 100-point scale.
Conversational AI Voice Coach:
- Grounded on the speaker's historical Voice Vault, allowing users to chat with an AI coach that tailors actionable speaking drills specifically to their longitudinal habits.

🏗️ How We Built It

Oratio was engineered as a modern full-stack web application designed for low latency and high reliability:

Frontend & App Framework: Next.js 16 (App Router) with React 19, Lucide Icons, and Vanilla Tailwind styling. Frameless, borderless card aesthetics designed for sleek dark mode.
Audio Processing:
- Browser-native Web Audio API and Web Speech API for zero-latency client-side DSP, real-time waveform buffer visualization, and live speech-to-text.
- Server-side acoustic heuristics for millisecond-level dead-air and tempo detection.
AI & Intelligence Engine:
- Gemini 2.5 Flash: Evaluates semantic rubrics, generates punchy executive summaries, extracts structural outlines, and powers the interactive voice coach.
- Deterministic Heuristic Engine: A resilient local acoustic analyzer that guarantees the platform functions 100% offline or during API rate limits.
Database & Persistence:
- Prisma ORM modeling speakers, speeches, temporal flaw annotations, and rubric score breakdowns.
- Hosted on a Supabase PostgreSQL cluster configured with transaction and session connection pooling.
Deployment: Hosted on Vercel Serverless Edge infrastructure.

🧗 Challenges We Faced

Prerender Suspense Boundaries in Next.js 16:
- Next.js 16 enforces strict suspense boundaries for client-side search parameter hooks (useSearchParams). Prerendering the /analyze studio initially triggered production build failures, which we resolved by isolating dynamic query readers inside dedicated <Suspense> wrappers.
Serverless Database Migration (SQLite to Supabase PostgreSQL):
- During initial deployment, local SQLite files (dev.db) failed inside Vercel's ephemeral, read-only lambda environment (ERROR 750680989).
- We migrated our data architecture to Supabase PostgreSQL, resolved IPv6-to-IPv4 routing across regional poolers (aws-0-ap-south-1.pooler.supabase.com:5432), and added safety guards across all server component queries so the application degrades gracefully under network latency.
Millisecond-Accurate Temporal Grounding:
- Aligning raw audio duration, word-level speech recognition segments, and LLM semantic critique required building a custom temporal alignment schema that synchronizes flaw timestamps with audio playback scrubbing.

🏆 Accomplishments That We're Proud Of

Sub-Second Feedback Loop: Users can record a 15-second voice memo and receive complete acoustic telemetry and timeline markers in under 2 seconds.
True Temporal Grounding: Instead of giving abstract advice, every critique is attached to an interactive timestamp button that jumps right to the audio flaw.
Zero-Downtime Hybrid Intelligence: The app works with or without external LLM keys—the DSP heuristics ensure users always get accurate delivery metrics.
Sleek, Frameless Aesthetic: A clean, distraction-free visual design that feels like a native desktop studio rather than a standard web form.

📚 What We Learned

Audio DSP on the Web: Working directly with the AudioContext and waveform buffers taught us the power and nuances of browser-native audio APIs.
Serverless PostgreSQL Optimization: Managing connection limits on serverless platforms reinforced the importance of connection poolers (PgBouncer/Supavisor) versus direct database connections.
Speech vs. Text Psychology: Spoken communication requires fundamentally different critique criteria than written prose; cadence variations and intentional pauses often communicate more than vocabulary complexity.

🚀 What's Next for Oratio

WhatsApp & Telegram Bot Integration: Send a voice note directly to a Telegram bot and immediately receive a 1-minute delivery report back.
Vocal Tone & Pitch Modulation Analysis: Adding pitch variance detection ($F_0$ fundamental frequency analysis) to detect monotone delivery or vocal fry.
Team Voice Vaults: Enabling engineering and product teams to benchmark asynchronous standup clarity and track speaking improvements over quarters.