Voice Intelligence — Dima Perkis
Voice Intelligence
Linguistic analysis of 1,438 recordings — 222 hours of meetings, coaching sessions, and solo notes. mlx-whisper transcription, LLM-powered speaker attribution, hedge rate tracking, acoustic feature extraction, and a RAG search layer across 1.95 million words.
What It Is
A quantified self-communication system. Every SuperWhisper dictation, Just Press Record capture, and Voice Memo gets transcribed locally via mlx-whisper, classified by type (meeting, coaching, solo note, interview), enriched with speaker attribution, and analyzed for linguistic markers — hedge rates, filler words, F0 pitch patterns, speaking pace.
What It Found
The “like” acceleration was the headline: usage climbed from 3.8 per 1,000 words to 27.9 over 16 months. Solo recordings showed half the hedge rate of meetings (7.5 vs 13.8), and conversations with my executive coach showed a consistent +39.6 Hz pitch elevation — measurable engagement.
The Stack
- mlx-whisper for local transcription (no cloud, no API costs)
- resemblyzer for voiceprint enrollment (16 speakers identified)
- ChromaDB for semantic search across 8,688 transcript chunks
- Claude for classification, speaker attribution, and enrichment
- Chart.js for a self-contained 10-panel visualization dashboard
Scale
| Metric | Value |
|---|---|
| Total recordings | 1,438 |
| Audio hours | 222 |
| Word count | 1.95M |
| Speakers enrolled | 16 |
| ChromaDB chunks | 8,688 |
| Date range | Dec 2024 — May 2026 |
Tools that helped
Python Claude