VoiceAI
Every sales call holds a deal or a missed one. VoiceAI reads the room — sentiment, signals, and the moment the pitch broke down.
About
An AI-powered voice analytics platform that transcribes sales calls and surfaces deep intelligence from every conversation. Upload or record any audio file — VoiceAI runs it through a two-stage pipeline: OpenAI Whisper for transcription, then GPT-4o for structured analysis. In Sales mode, it evaluates executive performance across 8 dimensions, predicts conversion probability, tracks talk ratios, and flags every buying signal, objection, and follow-up action buried in the call.
Core Features
Dual Analysis Modes
General mode extracts topics, sentiment, and action items. Sales mode goes deep — 8 executive dimensions, conversion probability, buying signals, objections, and talk ratios.
Whisper Transcription
OpenAI Whisper handles MP3, WAV, WebM, M4A, OGG, FLAC, AAC, and MP4. Speaker count and language are auto-detected from the audio.
Conversion Probability
GPT-4o synthesises executive performance, customer sentiment, and engagement level into a conversion probability score with the key factors that drove it.
Talk Ratio Analytics
Tracks who spoke when — executive vs customer talk time, silence gaps, interruptions, and the engagement balance across the full call.
Key Moment Detection
Every buying signal, objection, and question is timestamped and surfaced — so sales managers can jump straight to the moments that decided the deal.
Real-Time Recorder
Browser-based audio recording with a live Web Audio API waveform. No external tool needed — record, upload, and get analysis without leaving the app.
Voice Analysis Pipeline
End-to-end flow from audio input to structured intelligence report
Sales Mode
8-Dimension Executive Evaluation
A single GPT-4o call with a 3,500-token structured JSON schema evaluates every dimension of an executive's performance — scored and explained, not just labelled.
Customer Intelligence
Sentiment trajectory, engagement level, pain points, buying signals, objections, and unmet expectations — from the customer side of the call.
Conversion Probability
A confidence-weighted prediction score (0–100%) with the specific factors that pushed it higher or lower in this call.
Follow-Up Actions
GPT-4o generates prioritised next steps, open questions, and re-engagement hooks to act on before the next touchpoint.
What I Built
Designed a two-stage AI pipeline: Whisper handles speech-to-text for 8+ audio formats (MP3, WAV, WebM, M4A, OGG, FLAC, AAC, MP4) and then a GPT-4o prompt in strict JSON mode extracts structured intelligence — the pipeline produces the same predictable schema regardless of call length or audio quality
Built a Sales Analysis prompt that evaluates 8 executive dimensions (communication style, tone, listening, product knowledge, rapport, objection handling, confidence, closing skills) plus customer sentiment, engagement level, buying signals, objections, and pain points — all in a single GPT-4o call with a 3,500-token structured JSON schema
Implemented conversion probability prediction with a confidence score — GPT-4o synthesises all signal categories (executive performance, customer engagement, sentiment trajectory) into a probability estimate plus the specific factors that drove it up or down
Built a waveform-based real-time audio recorder in the browser using the Web Audio API frequency analyzer — users can record directly without leaving the app, with live visual feedback and a 50 MB file validation guard on both client and server
Architected the backend with Spring WebFlux async HTTP client configured with 120-second timeouts to handle long Whisper processing windows without blocking Tomcat threads — all external API calls are non-blocking reactive streams
Designed an aggregated Sales Analytics dashboard that pulls all SALES-type reports, calculates rolling averages across executive dimensions, and renders radar charts, talk-ratio pie charts, and trend bars using Recharts with Framer Motion entry animations
Tech Stack
Frontend
Backend
AI / Voice
Data
Explore the project