Talk

<p align="center"> <img src="docs/assets/icon_512.png" width="128" height="128" alt="Talk icon"> </p> <h1 align="center">Talk</h1> <p align="center"><strong>Open Typeless. Local Typeless. Typeless in your box.</strong></p>

A macOS menu bar voice input tool — hold a hotkey, speak, and your words are recognized, polished, and pasted into the active app. Your voice, straight to text. No cloud. No typing.

Download Talk v0.4.0 · 中文文档

The original algorithm and code are based on the generous contribution of @jiamingkong. We just wanted to see if we could build a typeless in ten minutes.

Features

• On-device inference — Powered by Apple Silicon MLX, no cloud dependency, privacy-first
• Dual ASR engines — Local MLX (Qwen3-ASR-0.6B-4bit) or Apple Speech Recognition, switchable in settings
• Text polishing — Qwen3-4B-Instruct, removes filler words, adds punctuation, smart formatting
• Auto hotword learning — Passively observes your edits after text injection, automatically learns ASR corrections (proper nouns, homophones, abbreviations) via LLM extraction
• Audio history — Every recording saved as AAC/M4A with full ASR context snapshot for replay and debugging
• Customizable prompts — Per-app prompt profiles, 3 polish intensity levels, or write your own system prompt
• Selection edit mode — Select text, speak a command ("fix the typo", "make it casual"), and it's done
• Floating status indicator — Always-on-top overlay showing recording/processing state with audio level meter
• Global hotkey — Customizable key recorder, Push-to-Talk / Toggle modes
• Audio device selection — Pick your input device, defaults to built-in microphone
• Auto-paste — Injects text via Accessibility API with CJK input method auto-switching
• Vocabulary learning — Automatic learning from edit history + manual entry, corrections injected into LLM context
• Idle memory management — Auto-unload models after inactivity, reload on demand

Performance

All inference runs on-device via Apple Silicon GPU. No network required after model download.

| Stage | Latency | Notes | |-------|---------|-------| | ASR (3-5s audio) | 0.07 - 0.18s | 17-51x faster than real-time | | LLM polish (short text) | 0.35 - 0.50s | ~30 chars input | | LLM polish (long text) | 1.1 - 1.2s | ~120 chars input | | Full pipeline | ~1s | ASR + LLM combined (models warm) | | ASR model load | 2s | Cold start, one-time | | LLM model load | 10s | Cold start, one-time — bottleneck |

Memory usage:

| State | RSS | |-------|-----| | ASR model loaded | ~1.6 GB | | Both models loaded | ~5.4 GB |

Full benchmark details and reproduction steps: docs/BENCHMARK.md

Run make benchmark to reproduce on your machine.

Compatibility

The pre-built DMG in Releases is built and tested on macOS 26.2 (Tahoe) with Apple Silicon. That's the only environment we have — our human overlords haven't blessed us with more test devices yet.

| | Tested | Should Work | Notes | |---|--------|------------|-------| | macOS 26.x (Tahoe) | ✅ | ✅ | Built & tested here | | macOS 15.x (Sequoia) | | Likely | Dependencies support macOS 14+ | | macOS 14.x (Sonoma) | | Maybe | Minimum required by MLX dependencies | | macOS 13 and below | | No | MLX framework requires macOS 14+ | | Intel Mac | | No | MLX is Apple Silicon only |

If you're on an older macOS version and encounter issues, try building from source — it might just work:

git clone https://github.com/platx-ai/Talk.git && cd Talk
make build && make run

If it doesn't, open an issue and tell us what broke. We'd love more test environments.

Requirements

• Apple Silicon (M1/M2/M3/M4) — required, no Intel support
• macOS 14.0+ (Sonoma) — minimum for MLX dependencies; pre-built DMG targets 26.2+
• 16 GB RAM recommended (8 GB works with lightweight model — coming soon)
• ~3 GB disk space (model files)
• Xcode 26.3+ (only for building from source)

Quick Start

# Clone the project
git clone https://github.com/platx-ai/Talk.git
cd Talk

# Full setup: resolve dependencies + download models
make setup

# Run
make run

Build

make build          # Debug build
make build-release  # Release build
make test           # Run unit tests
make benchmark      # Run performance benchmarks
make run            # Build and run
make clean          # Clean build artifacts
make resolve        # Resolve SPM dependencies only
make download-models # Download ML models from HuggingFace
make setup          # Full setup: resolve + download models
make lint           # Run SwiftLint (if installed)

Architecture

Record(AVAudioEngine) → ASR(Qwen3-ASR) → LLM Polish(Qwen3-4B) → Text Inject(Cmd+V)
       ↑                  0.1s               0.5s                     ↑
    CoreAudio                                                    Accessibility
  Device Selection                                                API Permission

Modules

| Module | Responsibility | |--------|---------------| | Audio/ | Recording engine, global hotkeys (CGEventTap), audio device management, text injection | | ASR/ | Speech recognition — MLX local (Qwen3-ASR) + Apple Speech | | LLM/ | Text polishing + hotword extraction (MLXLLM + Qwen3-4B-Instruct) | | Models/ | Data models (AppSettings, HotKeyCombo, HistoryItem, ASRContext) | | Data/ | History (JSON + M4A audio), vocabulary, edit observer | | UI/ | SwiftUI menu bar, settings panel, key recorder, floating indicator, history browser, flash capsule | | Utils/ | Logging system, Metal runtime validation |

Dependencies

All dependencies managed via Swift Package Manager, pinned to specific commits:

| Package | Source | Purpose | |---------|--------|---------| | mlx-swift | ml-explore/mlx-swift | MLX core array operations | | mlx-swift-lm | ml-explore/mlx-swift-lm | LLM inference framework | | mlx-audio-swift | platx-ai/mlx-audio-swift (fork) | Audio STT framework | | swift-huggingface | huggingface/swift-huggingface | Model downloading |

mlx-audio-swift uses the platx-ai fork to fix an upstream bug where MLXAudioCodecs is missing the MLXFast dependency.

Models

| Model | Size | Load Time | Memory | Purpose | |-------|------|-----------|--------|---------| | Qwen3-ASR-0.6B-4bit | ~400 MB | 2s | ~1.6 GB | Speech recognition | | Qwen3-4B-Instruct-2507-4bit | ~2.5 GB | 10s | ~4 GB | Text polishing |

Models are automatically downloaded from HuggingFace on first run to ~/.cache/huggingface/. Pre-download with make download-models.

Vocabulary & Auto Learning

Talk learns from your corrections in two ways:

Passive Edit Observation (v0.4.0)

After text is injected into the target app, Talk passively monitors the text field via Accessibility API. If you edit the injected text (e.g., fix a misrecognized word), Talk detects the change, extracts hotword corrections using a background LLM pass, and adds them to the vocabulary. A flash ⚡ capsule in the menu bar confirms when new corrections are learned. This works automatically — no manual steps needed.

Manual Correction

• History edit — Edit polished text in the history view. The system learns the correction automatically.
• Manual entry — Settings → Personal Vocabulary → Manage Vocabulary. Add original words and their corrected forms.
• Import/Export — JSON format via Manage Vocabulary.

The top learned corrections are injected into the LLM system prompt, so the model applies them automatically in future polishing.

Example: If ASR outputs "la laam" but you correct it to "LLM", future polishing will automatically apply this correction.

Audio History (v0.4.0)

Every voice input is saved as AAC/M4A (64kbps, ~80KB per 10s) alongside a context snapshot (hotword list, language, polish intensity, target app). This enables:

• Replay & debugging — Reproduce ASR issues with the exact audio that was processed
• Regression testing — Compare recognition quality across versions
• Automatic cleanup — Audio files are deleted when history entries are removed or expired

Toggle in Settings → Personal Vocabulary → "Save Audio History".

Permissions

On first launch, you need to grant:

1. Microphone — Required for recording. macOS will prompt automatically.
2. Input Monitoring — Required for the global hotkey. Enable Talk in System Settings → Privacy & Security → Input Monitoring.
3. Accessibility — Required for auto-pasting text into other apps. Enable Talk in System Settings → Privacy & Security → Accessibility.

If the global hotkey does not respond, check Input Monitoring first. After enabling it, quit and relaunch Talk so the hotkey listener can work reliably.

Development

# Open in Xcode
open Talk.xcodeproj

# Set your signing team: Xcode → Signing & Capabilities → Team
# Build & Run: ⌘R

Testing

make test       # Unit tests
make benchmark  # Performance benchmarks (ASR/LLM load, inference, pipeline, memory)

All changes require tests. Bugs require regression tests before fixing. See CLAUDE.md for testing rules.

Code Signing

DEVELOPMENT_TEAM is left empty in the project. Each developer sets their own signing team in Xcode. CLI builds use ad-hoc signing.

Roadmap

See ROADMAP.md for the full product roadmap.

Next up

• Custom lightweight polish model (0.5-1.5B) — < 1s load, < 1 GB memory