Echo

Private voice. Local intelligence.

Live case study · Product demo · Benchmark methodology · Architecture

Important

This is a public engineering case study, not Echo’s product source repository. Echo remains a private pre-release product. This repository contains reviewed documentation and non-sensitive visual evidence only.

What Echo is

Echo is a standalone, local-first macOS voice workspace built around two workflows:

• Cross-application dictation: capture speech, transcribe locally, apply deterministic vocabulary and spoken-command rules, then insert into the active app with a safe clipboard fallback.
• Evidence-grounded meeting intelligence: record or import meetings, create a local transcript, map candidate facts, validate them against transcript evidence, and reduce the validated set into structured notes.

Echo also provides one-way, graph-ready Obsidian export for meetings and user-confirmed entities. Echo remains the canonical local store, and source audio is not written to the vault.

The engineering problem

Most voice products optimize the model call. Echo treats the complete interaction as the product:

1. Can capture remain reliable across macOS permission and focus changes?
2. Can recognition feel immediate without sending speech to an API?
3. Can private vocabulary be corrected without claiming model-weight fine-tuning?
4. Can meeting notes retain exact evidence rather than relying on plausible prose?
5. Can text reach the intended app without accidentally sending a message?

System architecture

flowchart LR
    subgraph D["Private dictation"]
        A["Microphone capture"] --> B["VAD + silence gates"]
        B --> C["Parakeet TDT 0.6B v3\nCore ML"]
        C --> E["Deterministic vocabulary\n+ spoken commands"]
        E --> F{"Target still safe?"}
        F -->|"Yes"| G["Insert into active app"]
        F -->|"No"| H["Safe clipboard delivery"]
    end

    subgraph M["Meeting intelligence"]
        I["Live capture or audio import"] --> J["Local transcript"]
        J --> K["Map candidate evidence"]
        K --> L["Validate exact support"]
        L --> N["Local Qwen reduction"]
        N --> O["Summary · decisions · actions"]
    end

Loading

The production dictation route is Parakeet TDT 0.6B v3 through Core ML. Meeting intelligence uses a local Qwen3.5 MLX 4-bit route combined with deterministic evidence validation. Whisper large-v3-turbo remains a meeting transcription benchmark and fallback; it is not the production dictation route.

See Architecture for the component boundaries and sanitized pseudocode.

Selected evidence

Result	What it means	Boundary
4.74% WER	Mean raw WER across 11 generated speech fixtures	Generated regression suite, not a broad real-world claim
≈85 ms	Average model-only time on the generated production-route suite	Excludes capture, load, delivery, and UI overhead
0 / 2 false positives	Silence and low-background-noise fixtures returned empty transcripts	Two generated non-speech fixtures only
3.38% WER	Echo Parakeet result on a 100-clip common public-audio track	Same sampled clips and normalization across compared local routes
6.19%*	Reconstruction over 634 released AA-WER v2 public samples	Not an official AA-WER score or rank
5 / 5 and 9 / 9	Grounded meeting-output contracts passed	Five synthetic domains and one corrected private financial-call reference

The complete protocol, hardware context, and limitations are documented in Benchmark Methodology.

What failed—and what changed

Qwen3-ASR 0.6B INT8 did not replace Parakeet

On Echo’s generated suite it produced 13.19% WER with approximately 2.94 seconds average model time, including severe weakness on technical vocabulary. The challenger was rejected rather than hidden behind a blended score.

Direct structured meeting generation was not reliable enough

Smaller-model experiments could produce convincing prose while violating the grounded-output contract. Echo moved evidence validation in front of the final reduction: unsupported candidate facts are removed before they can become notes.

Recognition and delivery could not be one action

A correct transcript can still be dangerous if focus changes. Echo separates recognition, deterministic rewriting, target validation, and delivery. When direct insertion is unavailable, the text is preserved through a safe-copy fallback.

See Engineering Decisions for the longer decision log.

Evidence boundaries

This case study deliberately separates four kinds of claims:

• Measured evidence: results reproduced by Echo’s committed benchmark protocols.
• Implemented behavior: capabilities present in the private application and covered by product tests.
• Published context: external model or cloud results shown only for orientation.
• Future work: release, validation, or product improvements that are not presented as complete.

Echo does not claim model-weight fine-tuning. Vocabulary correction and spoken-command processing are deterministic application logic.

Technology

Swift · SwiftUI · AppKit · Core ML · FluidAudio · Parakeet · MLX Swift · Qwen3.5 · local audio capture · deterministic evidence validation · macOS Accessibility APIs

Repository boundary

What is public here:

• Product and architecture documentation
• Sanitized pseudocode
• Benchmark methodology, selected results, and limitations
• Demo imagery and links to the public case study
• Engineering decisions, including rejected approaches

What remains private:

• Product source code and Git history
• Proprietary workflow and release logic
• Credentials, signing identities, and notarization records
• Private transcripts, tester data, and unreleased features
• Model weights and private benchmark references

This repository is not an SDK, binary distribution, or open-source release. All product rights are reserved.

More detail

• Architecture and data flow
• Benchmark methodology and claim ledger
• Engineering decisions and rejected approaches
• Privacy and public-release boundary