NEW v1.2.0 100% On-Device macOS AI Voice Assistant · 5 Refinement Levels & Instant Local Cleanup
Looking for Remote Buddy? Visit remote.zfdang.com →
Voice Buddy Mascot
voice buddy. Voice Assistant · macOS
100% On-Device Inference · Zero Cloud API Costs · Multilingual

Think it, speak it,
it's written.

Never break your flow to type. Hold a hotkey anywhere, speak naturally, and release. On-device acoustic models and local LLMs filter filler words, fix grammar, and polish punctuation—typing clean text directly at your cursor.

💻 Designed for Apple Silicon (M-Series) · macOS 15.0+ View v1.2.0 Release Notes →
Listening & Streaming · Real-Time Unified Memory
Dynamic Streaming HUD
Voice Buddy Floating Dictation HUD
🛡️ Audio Stays Local 🧠 On-Device LLM Polish ⚡ Sub-Second Response
✳ Keep your thought train moving. LOCAL-FIRST AI v1.2.0
✦ 100% On-Device Inference ✦ Zero Cloud API Costs ✦ Full Offline Availability ✦ Verbal Filler Filtering ✦ Password & Clipboard Guard
00 / PRIVACY & SECURITY FIRST

Your voice belongs to you, and only you.

All acoustic transcription and LLM refinement happen inside Apple Silicon unified memory—zero internet required, eliminating leaks from the hardware up.

🛡️

Zero Cloud Transmission

Microphone streams are processed entirely in local RAM. Nothing is ever transmitted to remote cloud servers. Pull your Ethernet cable, toggle Airplane Mode—it keeps dictating.

🔥

Zero Audio Storage

Audio buffers are cleared immediately from memory after processing. No audio files, recordings, or cache files are ever written to your local disk.

💎

Zero API Cost or Quota Anxiety

No user accounts to register, no third-party API keys to configure, and zero per-token billing anxieties. Download and start speaking immediately.

🔒

Secure Input & Transient Clipboard

Dictation is automatically intercepted on secure password fields to prevent credential leaks. Injections use transient clipboard markers to evade third-party clipboard scrapers.

Read our full Privacy & Security Commitments on GitHub ↗
01 / DUAL-STAGE INTELLIGENCE

Speech-to-text plus LLM refinement, side-by-side.

Spoken language naturally includes filler words, false starts, and repetitions. Voice Buddy relays acoustic transcription directly to local LLMs, refining spoken phrasing into polished written prose.

Try 5 Text Refinement Levels:
Raw Acoustic Transcript (ASR) Qwen3-ASR / Whisper Turbo / SenseVoice
“um hey David so regarding next week's release schedule let's have a quick sync at two pm and QA team please have the test cases ready that's it”
Direct acoustic output · Includes verbal fillers & missing punctuation
AI Refined & Polished Text Deterministic Local Cleanup · Light Polish (Level 1 · Default)
“Hey David, regarding next week's release schedule, let's have a quick sync at 2 PM, and QA team please have the test cases ready.”
✓ Instant local rule cleanup · Removes hesitations and normalizes spacing · Zero model memory overhead v1.2.0 Tested in Results Center
✨

5 Refinement Levels in Settings: Tailor the refinement depth anytime in Settings → General → Text Refinement: Original (Raw), Light Polish (Level 1 · Local Rules), Moderate (Level 2), Deep Polish (Level 3), or Structured Summary (Level 4). Compare both versions side-by-side in Results anytime!

02 / THREE-STEP WORKFLOW

Effortless typing in three natural steps.

Direct injection into the active cursor without window switching.

STEP 01

Place Your Cursor

Slack, Notion, VS Code, Terminal, WhatsApp, or Google Docs. Wherever your cursor rests, words appear directly.

STEP 02 Fn / 🌐

Hold and Speak

Press and hold Fn / Globe or ^⌥Space to speak freely. Supports both "Hold to Talk" and "Toggle On/Off" modes.

STEP 03

Release to Type

The moment you release, acoustic recognition and LLM refinement complete in milliseconds—typing words directly at your cursor.

03 / ON-DEVICE MODELS

Switch Engines Freely, Tailored for Your Mac.

Five leading on-device acoustic models optimized for diverse hardware specs and language workloads.

Whisper Large v3 Turbo

CoreML · 980 MB

⚡ 300~600ms

High-precision transcription for 90+ global languages, translation, and pure English.

  • Comprehensive coverage of major international languages
  • Accelerated via Apple Neural Engine (ANE) hardware
  • Optimized compilation on first run for ultra-fast standby calls

SenseVoice Small

CoreML · 230 MB

⚡⚡ 30~80ms Ultra-Fast

Sub-100ms response with minimal RAM footprint for everyday chat.

  • Alibaba FunASR classic, instant on-screen typing on key release
  • Low memory overhead, runs effortlessly on 8GB / 16GB Macs
  • Ideal for instant messaging, quick replies, and voice notes

Qwen 3 ASR 0.6B

Recommended

MLX 8-bit · 960 MB · ~1.0 GB Unified Memory

⚡ 200~400ms

Lightweight & fast · Best daily balance for programming, technical terms, and bilingual dictation.

  • Tailored for software engineering and technical communication
  • Minimal memory footprint, seamless on 8GB / 16GB Apple Silicon
  • Millisecond low-latency streaming with accurate code keyword recognition

Qwen 3 ASR 1.7B

Flagship Precision

MLX 8-bit · 1.7 GB · ~1.8 GB Unified Memory

⚡ 500~900ms

Flagship accuracy · Built for long complex sentences, technical seminars, and deep code reviews.

  • Highest accuracy for complex phrasing and academic terminology
  • Gold standard for lengthy technical lectures and code analysis
  • Deeply integrated with Apple Silicon Metal GPU unified memory

Apple Speech

System Native · 0 MB

⚡ 100~200ms

macOS built-in dictation service, zero disk space, zero download needed.

  • Ready to use out of the box with zero model downloads
  • Native system-level integration, ultra-lightweight
  • Energy efficient with 24/7 low-power standby
🧠

On-Device Language Models: Paired with on-device Qwen LLMs (Qwen 3.5 0.8B INT5/INT8, Qwen 3 1.7B, Qwen 3 4B), providing private text restructuring and punctuation cleanup across all acoustic engines.

Remote Buddy Logo
SISTER PROJECT · THE PERFECT PAIR

Use with Remote Buddy

Transform Xiaomi Bluetooth voice remotes or Chromecast remotes into wireless virtual microphones and hotkey controllers. Lean back in your chair and trigger Voice Buddy with a single thumb click.

Voice Buddy Mascot
READY TO GET STARTED?

Let your thoughts flow freely.

100% on-device inference · Absolute privacy · Zero cloud costs · Download and try today.

Designed for Apple Silicon (M1+) · macOS 15.0+