Overhear

Local-only live answer assistant for macOS. It listens to a meeting, interview, or class, optionally watches your screen, and answers questions in real time — fully offline. Nothing leaves your Mac: no cloud, no account, no telemetry. The only network calls are to a local Ollama server on 127.0.0.1.

See ABOUT.md for a feature-level overview.

Download

Download the latest release →

Grab Overhear_vX.Y.Z.dmg, drag Overhear to Applications. The app is ad-hoc signed (not notarized), so on first launch right-click the app → Open → Open (or System Settings ▸ Privacy & Security ▸ Open Anyway). You'll also need Ollama running locally. Requires macOS 26+, Apple Silicon.

Prefer to build it yourself? See Build & run below.

How it works

Piece	Technology
Speech-to-text	Apple SpeechAnalyzer / SpeechTranscriber (on-device, Neural Engine). en_US assets auto-download once from Apple.
Language model	Ollama on 127.0.0.1:11434 — one multimodal model does classification, text answers, and vision
Microphone	AVCaptureSession, pinned to the built-in mic
System audio + screenshots	ScreenCaptureKit
On-screen text	Apple Vision framework OCR
Echo / bleed control	Vendored Speex acoustic echo canceller (ThirdParty/speexdsp) + a text-overlap bleed filter
UI	SwiftUI in a non-activating NSPanel overlay, excluded from screen shares

Speaker split is by source, not diarization: your built-in mic is You, system audio is Them.

Requirements

• macOS 26+, Apple Silicon
• Ollama: run ollama serve, then pull a model, e.g. ollama pull qwen2.5:7b (any Ollama model works; a vision-capable model is needed only for the "Use screen" image mode — OCR mode works with any model)
• XcodeGen: brew install xcodegen

Build & run (dev)

xcodegen generate        # regenerate the .xcodeproj after adding/removing files
open Overhear.xcodeproj   # then ⌘R

The Xcode project is generated — edit project.yml, never the .xcodeproj (it's git-ignored).

Signing

The app uses a free Apple Development identity (own-machine use). In Xcode: target Overhear → Signing & Capabilities → Automatically manage signing → set Team to your Apple ID.

Entitlements (Resources/Overhear.entitlements):

• com.apple.security.device.audio-input — microphone access
• com.apple.security.cs.disable-library-validation — allows the vendored Speex C library to load under the hardened runtime

If the bundle id com.davidwang7723.Overhear is taken, change PRODUCT_BUNDLE_IDENTIFIER in project.yml and re-run xcodegen generate.

Package a signed DMG

./build.sh        # prompts for a version, e.g. 1.0.1

build.sh regenerates the project, builds Release, signs with the first Apple Development identity in your keychain, and writes Overhear_vX.Y.Z.dmg to your Desktop. Override the identity with OVERHEAR_SIGN_ID=<sha1-or-name> ./build.sh.

An Apple Development signature only runs on your own machines. To share the app with others, sign with a Developer ID certificate and notarize (notarytool).

First run

Grant Microphone and Screen & System Audio Recording when prompted (Screen Recording is what captures the other party's audio). For best accuracy during live calls, use headphones — with speakers, the other side's audio physically re-enters your mic, and while the echo canceller and bleed filter reduce this, they can't fully remove it during simultaneous speech.

Layout

Sources/
  App/            AppState (pipeline), OverhearApp (menu-bar agent)
  Audio/          mic + system-audio capture, Speex echo canceller, VAD
  Transcription/  SpeechAnalyzer wrapper
  LLM/            Ollama client + answer engine (prompt lives here)
  Store/          transcript + session persistence
  UI/             overlay panel, transcript/answer views, settings
ThirdParty/
  speexdsp/       vendored Speex DSP (BSD-3, see ThirdParty/speexdsp/COPYING)

License

See LICENSE. Bundled Speex DSP is BSD-3-Clause; its copyright notices are retained in each source file and in ThirdParty/speexdsp/COPYING.