01 / Private pre-release product
Echo
A clean, local-first macOS voice workspace for dictation, meetings, evidence-grounded notes, and structured memory.
- Swift
- SwiftUI
- Core ML
- FluidAudio
- MLX Swift
- Qwen
Deterministic vocabulary correction reduced delivered-text WER from 4.74% to 4.09% on Echo’s retained 13-fixture generated regression suite.
Voice software is useful only when capture is dependable, private speech stays private, model behavior is measurable, and final text reaches the right application safely. Echo treats those concerns as one product rather than a transcription demo.
Role
What I owned
Product direction, Swift architecture, local-model integration, evaluation, reliability, and private release engineering.
Original contribution
- Designed a clean Swift package with separate core, speech, macOS delivery, local intelligence, app, benchmark, and test targets.
- Integrated explicit local model preparation so ordinary dictation never triggers an unexpected download.
- Built deterministic vocabulary and formatting stages that preserve raw transcript evidence.
- Connected meeting transcripts to evidence-grounded summaries, decisions, actions, and graph-ready Obsidian knowledge.
- Added signing, packaging, restart-soak, release-acceptance, and private-tester DMG gates.
System / Architecture
How it works
- 01
EchoCore defines dictation, meetings, evidence, memory, benchmarks, and Obsidian contracts.
- 02
EchoSpeech runs local Parakeet recognition through FluidAudio and evaluates challenger routes.
- 03
EchoIntelligence runs local Qwen meeting intelligence through MLX Swift.
- 04
EchoMac owns permissions, capture, hotkeys, status, and safe text delivery.
Evidence / Protocol included
Measured results
Generated-suite delivered WER
4.09% after deterministic correction
Model: NVIDIA Parakeet TDT 0.6B v3 through FluidAudio 0.12.6/Core ML. Hardware: Apple-silicon Mac. Dataset: 13 generated fixtures—11 speech plus 2 silence/noise—covering English, Spanish, code-switching, numbers, and technical terms. Metric: corpus word error rate after final text correction. Timing boundary: prepared-model inference only; correction adds negligible measured cost. Echo-measured regression evidence, not an independent or published leaderboard score; generated speech is easier than diverse live microphones.
Technical-term fixture WER
7.14% → 0%
Same model, runtime, and Apple-silicon hardware. Dataset: one retained generated technical-language fixture. Metric: word error rate before and after deterministic delivered-text correction; timing excludes model preparation and UI delivery. Echo-measured narrow regression: a known ‘Quen’ → ‘Qwen’ rule was corrected, so this does not establish general acoustic or vocabulary accuracy.
Public AA-WER subset
3.25% VoxPopuli · 9.13% Earnings22
Model: NVIDIA Parakeet TDT 0.6B v3 through the production Core ML route. Hardware: Apple-silicon Mac. Dataset: all 634 publicly released AA-WER v2 VoxPopuli and Earnings22 samples, 3.89 hours. Metric: normalized corpus WER; timing boundary covers prepared-model transcription, not download or UI delivery. Echo-measured on a published public subset, not independently reproduced or an official AA-WER score: AgentTalk and the complete official normalizer were unavailable.
Limits
What this does not prove
- Source code and underlying evidence files are not public.
- Current neural weights are upstream models; Echo enhances the product and inference pipeline rather than claiming newly trained foundation-model weights.
- Generated speech is easier than diverse microphones, accents, far-field rooms, and interruptions.
- Public distribution and final live-dictation acceptance remain pre-release work.
Inspect the work
See the artifact, not just the claim.