MacWispr

Voice dictation for macOS
On-device Local STT — Qwen (En + Asian) on GPU or Parakeet v3 (En + EU) on Neural Engine — or bring your own OpenAI / ElevenLabs key.
A free, open-source alternative to Wispr Flow.

Hold a hotkey, speak, release — text appears in whatever app you're typing in.

Website

Marketing site: fuckwisprflow.com (MacWispr vs cloud)
Product notes: vasanthsreeram.github.io/macwispr

Easy install (stable)

Grab MacWispr 1.2.3 (stable / Latest) DMG / .app zip from Releases · Latest
Unzip / open DMG and drag MacWispr.app into Applications
Open it (right-click → Open the first time if macOS warns)
Grant Microphone + Accessibility
Hold ⌥Space, speak, release
Optional: Settings → Transcription to add OpenAI / ElevenLabs keys (BYOK)

Beta (1.2.4-beta.1) — bundled on-device polish

Opt-in pre-release with Qwen3.5-0.8B polish bundled (~1.2 GB download). Does not replace stable Latest or Sparkle. Developer ID + notarized.

Release: v1.2.4-beta.1
DMG: MacWispr-1.2.4-beta.1-macos-arm64.dmg
Also includes multi-window / history fixes (#14 / #15) and content-free polish telemetry (opt-in)
After install: open from Applications (not a repo dist/ copy) · Settings → Post-Processing → Local LLM for polish
Scope / clean-install notes: docs/context/RELEASE_1.2.4_BETA.md

One-command build & install (from source)

git clone https://github.com/vasanthsreeram/macwispr.git
cd macwispr
./scripts/install.sh
open -a MacWispr

Build the app bundle only:

./scripts/build-app.sh
open dist/MacWispr.app

Developer quick run (no .app bundle):

swift build -c release
.build/release/MacWispr

Benchmark latency + accuracy (WER):

./bench.sh
# WER on LibriSpeech subset:
uv run --with 'git+https://github.com/Blaizzy/mlx-audio.git' --with soundfile \
  python bench/bench_wer.py --max-files 50 --models default

Features

Hold-to-dictate — Hold Option+Space, speak, release to transcribe and insert (toggle mode available)
System-wide insertion — Pastes into Slack, VS Code, browser, terminal, anywhere
On-device ASR — pick by language coverage in Settings or the dashboard Local chip:
- Qwen 0.6B / 1.7B (En + Asian) — GPU (MLX)
- Parakeet v3 (En + EU) — Neural Engine (Core ML)
RAM-aware default — Qwen 1.7B when Mac has >16 GB RAM; else Qwen 0.6B (override anytime)
BYOK cloud STT — Optional OpenAI (gpt-4o-mini-transcribe) or ElevenLabs (scribe_v2); keys in Keychain only
Transcript polish — Off, local Qwen3.5 polish SFT (bundled in 1.2.4-beta), or OpenAI BYOK; polish runs before paste
Menu bar app — Live status (red mic + timer while listening; Done shows STT latency)
Listening banner — Optional floating “Listening / Done” under the menu bar (not a fake Dynamic Island)
Configurable chimes — Per-event system sound + volume in Settings (start / stop / done / error)
Weekly Time Saved dashboard — Word count + estimated typing time saved (baseline up to 200 WPM)
Sparkle auto-updates — Check for Updates via appcast on fuckwisprflow.com (stable line)
Opt-in telemetry — Anonymous, content-free reliability events (off by default; see PRIVACY.md)
Cleanup — Optional auto-capitalize; fillers/lists via polish model (no hardcoded filler word list)
Dev capture (beta) — Optional local WAV + raw/polished text for debugging (never telemetried)
52 languages — Auto-detect or pin a language (Qwen / cloud; Parakeet is multilingual EU set)
Transcription history — Browse and copy past results (persisted locally)
Multiple insertion modes — Clipboard paste, simulated typing, or both
Custom vocabulary — Names/jargon bias for Qwen + cloud (not applied to Parakeet)

Weekly Time Saved dashboard

MacWispr tracks every dictation and shows how much typing time you avoided.

Metric	What it means
Words	Word count for the last 7 days
Time saved	Estimated typing time at your baseline WPM minus time spent speaking
Spoken	Total audio captured
Words/day chart	Daily breakdown for the current week

Default typing baseline is 40 WPM (adjust in Settings → Dashboard). History lives only on your Mac under Application Support.

Menu bar also shows this week’s words + time saved at a glance:

Requirements


OS	macOS 14.0+ (Sonoma or later)
Chip	Apple Silicon (M1 / M2 / M3 / M4 / M5)
Xcode	Full Xcode.app is required — Command Line Tools alone are not enough. MLX needs the Metal shader compiler (`metal`). With only CLT, `swift build` can succeed but the app fails at runtime with `Failed to load the default metallib`. Install Xcode, then `sudo xcode-select -s /Applications/Xcode.app/Contents/Developer`.
Disk	~500 MB–1.5 GB+ for local models on first use (Qwen and/or Parakeet); cloud BYOK needs no local ASR download
Permissions	Microphone + Accessibility

Preflight (also run automatically by ./bench.sh, ./scripts/install.sh, and ./scripts/build-app.sh):

./scripts/preflight-xcode.sh
# equivalent check:
xcrun -sdk macosx metal --version

Usage

Launch the app — local model auto-downloads on first run (cached under ~/Library/Caches/)
Grant Microphone and Accessibility when prompted
Hold Option+Space, speak, release — text appears in the focused field
Open the main window for the Dashboard, history, and Settings
For cloud STT: Settings → Transcription → Provider → OpenAI or ElevenLabs → save your API key

Benchmark

MacWispr includes harnesses for latency and accuracy (WER) against its default Qwen3-ASR stack and the engines FluidVoice-class apps use (Parakeet TDT v2/v3, Nemotron 3.5). Weights are official mlx-community ports via mlx-audio, plus optional FluidAudio CoreML.

# Latency — Qwen (MLX) + Parakeet v3 (CoreML via speech-swift)
./bench.sh

# Latency — cross-engine (adds Nemotron / Parakeet MLX)
./bench/bench_compare.sh

# Latency — mlx-community only
uv run --with 'git+https://github.com/Blaizzy/mlx-audio.git' \
  python bench/bench_mlx_audio.py bench/clips/speech_12s_16k.wav --models default

# Accuracy (WER) — LibriSpeech test-clean subset
uv run --with 'git+https://github.com/Blaizzy/mlx-audio.git' --with soundfile \
  python bench/bench_wer.py --max-files 50 --models default
# full set: --max-files 0    | also Qwen 1.7B: --models all

Details and flags: bench/README.md. Snapshot of the latest local run: bench/results/SUMMARY.md.

Method

Kind	Method
Latency	load once → warmup → 3 timed runs → best latency; RTF = latency ÷ audio duration
Accuracy	LibriSpeech test-clean subset; aggregate WER = total edits ÷ total ref words (lowercase, strip punctuation)

RTF under 1.0 means faster than realtime.

At a glance (Apple M5, 32 GB)

Engine	Latency (~9.6 s clip)	RTF	LibriSpeech WER (50 utts)	Role
Parakeet TDT v2 MLX	0.081s	0.009	0.72%	Fastest + best clean-EN WER
Parakeet TDT v3 MLX	0.091s	0.009	0.92%	Multilingual batch (FluidVoice-class)
Qwen3-ASR 0.6B 8-bit MLX	~0.4s (RTF 0.040 on LS)	0.040	0.92%	App option — 52 langs; default on ≤16 GB
Nemotron 3.5 0.6B 8-bit MLX	0.409s	0.043	1.64%	Streaming / multi-locale
Parakeet v3 CoreML (FluidAudio CLI)	~0.18s wall	~0.019	—	ANE path; warm process wall time

Speed (lower latency better)          Accuracy on clean EN (lower WER better)
Parakeet v2   █ 0.08s                 Parakeet v2   █ 0.72%
Parakeet v3   █ 0.09s                 Qwen 0.6B     ██ 0.92%
Nemotron      █████ 0.41s             Parakeet v3   ██ 0.92%
Qwen 0.6B     █████ ~0.4s             Nemotron      ████ 1.64%

Takeaway: On clean English, all four open engines are under ~2% WER on our 50-utt set. Parakeet wins raw speed (~4× Qwen RTF); Qwen ties Parakeet v3 on WER while covering more languages and remaining the product default.

Latency detail

Same speech clip (~9.6 s, 16 kHz mono, bench/clips/speech_12s_16k.wav). Measured 2026-07-11/12 on Apple M5 32 GB.

xychart-beta
    title "Inference latency on ~9.6s speech — lower is better"
    x-axis ["Parakeet v2 MLX", "Parakeet v3 MLX", "Parakeet v3 CoreML", "Nemotron 8bit", "Qwen 0.6B"]
    y-axis "Seconds" 0 --> 0.5
    bar [0.081, 0.091, 0.18, 0.409, 0.40]

Model	Backend	Latency	RTF	Speed	Notes
Parakeet TDT 0.6B v2	MLX (`mlx-community`)	0.081s	0.009	~118×	English-only; fastest pure MLX
Parakeet TDT 0.6B v3	MLX (`mlx-community`)	0.091s	0.009	~105×	Multilingual (25 EU langs)
Parakeet TDT 0.6B v3	CoreML / ANE (FluidAudio CLI)	~0.18s wall	~0.019	~53×	Warm process wall; cold first load ~208s (one-time compile)
Parakeet TDT 0.6B v2	CoreML / ANE (FluidAudio CLI)	~0.20s wall	~0.021	~48×	English-only CoreML
Nemotron 3.5 streaming 0.6B	MLX-8bit (`mlx-community`)	0.409s	0.043	~23×	Cache-aware streaming; ~40 locales
Qwen3-ASR 0.6B 8-bit	MLX	~0.4s (RTF 0.040 on LibriSpeech)	0.040	~25×	Default on ≤16 GB Macs; 52 languages
Qwen3-ASR 0.6B 4-bit	MLX	~0.60s	~0.06	~17×	Smaller / slightly worse quality
Qwen3-ASR 1.7B 8-bit	MLX	~0.77–1.78s	~0.08–0.18	~5–13×	Default on >16 GB Macs; highest local Qwen accuracy
Parakeet v3 (En + EU) in-app	Core ML / ANE	fast (ANE)	—	—	Dashboard Local chip · Settings → Model
ElevenLabs Scribe v2 (BYOK)	Cloud	~1.6–2.3s	network	—	Optional cloud STT

Accuracy detail (WER)

LibriSpeech test-clean subset: 50 utterances, ~376 s audio, 977 reference words.
Normalizer: lowercase + strip punctuation. Measured 2026-07-12 with bench/bench_wer.py.

Model	Aggregate WER ↓	Mean WER	Errors / words	Subs / Ins / Del	RTF
Parakeet TDT 0.6B v2 (English)	0.72%	1.71%	7 / 977	5 / 0 / 2	0.011
Qwen3-ASR 0.6B 8-bit	0.92%	2.14%	9 / 977	8 / 0 / 1	0.040
Parakeet TDT 0.6B v3 (multilingual)	0.92%	2.18%	9 / 977	8 / 0 / 1	0.011
Nemotron 3.5 streaming 0.6B 8-bit	1.64%	3.22%	16 / 977	13 / 1 / 2	0.043

Parakeet v2   ██░░░░░░░░  0.72%   ← best accuracy (this EN subset)
Qwen 0.6B     ███░░░░░░░  0.92%   ← MacWispr default
Parakeet v3   ███░░░░░░░  0.92%   ← same error count as Qwen, ~4× faster
Nemotron      █████░░░░░  1.64%   ← still strong; streaming-oriented

Caveat: 50 utterances is a local iteration set (first files in the 1–30 s duration window), not a full leaderboard. For publication-grade numbers: --max-files 0 (full 2620-utt test-clean). External full-set tables: soniqo.audio/benchmarks.

How to read this

Question	Answer on M5
Fastest open ASR?	Parakeet TDT (~100× realtime; ~0.08–0.09 s on ~10 s speech)
Most accurate on clean EN?	Parakeet v2 (0.72% WER); Qwen 0.6B and Parakeet v3 tie at 0.92%
v2 vs v3?	v2 English-only, best EN WER; v3 multilingual at the same WER as Qwen 0.6B here
Is Nemotron competitive?	Yes for streaming size/speed; 1.64% WER vs ~0.9% for Qwen/Parakeet on this set
Why ship Qwen as default?	Matches Parakeet v3 accuracy on this EN set, 52 languages, strong dictation defaults; still ~25× realtime
FluidVoice “Fluid Intelligence”?	Closed ~3.5 GB local enhancer (not open weights). MacWispr polish: Qwen3-0.6B-Chat CoreML or OpenAI BYOK

Qwen3 family only (in-app size ladder)

Earlier M5 family ladder (~10 s audio, speech-swift MLX, best of 3 after Metal warmup):

Model	Latency	RTF	Verdict
0.6B MLX-4bit	0.60s	0.060	Fast / smaller
0.6B MLX-8bit (app default)	~0.4s (see mlx-audio RTF 0.040 above)	~0.04	Default quality/speed balance
1.7B MLX-4bit	~1.20s	0.120	Higher accuracy, slower
1.7B MLX-8bit	~0.77–1.78s	~0.08–0.18	Best local Qwen accuracy
HF PyTorch 0.6B (MPS)	39s	1.31	Not viable for dictation

Optimizations in MacWispr

16 kHz capture — matches model input, no wasted resampling
Metal warmup on load — first dictation isn't a cold-start penalty
Dynamic max tokens — scales with utterance length instead of always using 448
Model stays loaded — no reload between dictations

Run the harnesses on your own Mac — absolute ms and WER vary by chip, power mode, thermal state, and utterance set.

Architecture

Coding agents: see AGENTS.md and docs/context/ARCHITECTURE.md.

Sources/
  MacWisprApp.swift          App entry (menu bar + window)
  AppState.swift             State, phases, hotkey wiring, providers, post-processing
  ASRModelSize.swift         On-device models: Qwen (En+Asian) + Parakeet v3 (En+EU)
  TranscriptionEngine.swift  Local Qwen3ASR (MLX) + ParakeetASR (Core ML)
  DashboardView.swift        Time Saved + Local model quick-switch chip
  CloudSTTClient.swift       OpenAI + ElevenLabs STT / OpenAI polish
  KeychainStore.swift        BYOK API keys (Keychain only)
  TranscriptionProvider.swift  Local / OpenAI / ElevenLabs + polish modes
  TextPolisher.swift         Optional on-device LLM polish
  AudioRecorder.swift        Mic capture, resample to 16 kHz
  HotkeyManager.swift        Global Option+Space (tap + Carbon + monitors)
  ListeningHUDController.swift  Optional banner under menu bar (Listening / Done + latency)
  FeedbackSounds.swift       Configurable system chimes + volume
  FailureBannerController.swift / OnboardingView.swift
  Telemetry.swift            Opt-in PostHog batch client
  TextInserter.swift         Clipboard paste or simulated typing
  SparkleUpdater.swift       Check for Updates
  SettingsView.swift         Simplified General / Transcription / Hotkeys / About

scripts/
  build-app.sh               Package MacWispr.app + Sparkle + sign
  install.sh                 Build and install to /Applications
  release.sh                 Tag + publish GitHub Release
  sign-and-notarize.sh       Developer ID + optional notary

docs/
  index.html                 Product website (GitHub Pages)
  SPARKLE.md                 Auto-update keys, sign_update, appcast deploy
  context/                   Agent-oriented architecture / issues / signing
  assets/                    Logo + README images

website/
  appcast.xml                Sparkle feed (deploy to fuckwisprflow.com)

Dependencies

soniqo/speech-swift — Qwen3-ASR (MLX), ParakeetASR (Core ML), SpeechVAD, AudioCommon
sparkle-project/Sparkle — in-app Check for Updates… (appcast on fuckwisprflow.com). Setup and release signing: docs/SPARKLE.md.

Troubleshooting

swift package resolve hangs? Manually fetch SpeechCore:

curl -L -o /tmp/SpeechCore.xcframework.zip \
  "https://github.com/soniqo/speech-core/releases/download/v0.0.3/SpeechCore.xcframework.zip"
# Extract to .build/checkouts/speech-swift/SpeechCore.xcframework/

Benchmark needs ffmpeg?

brew install ffmpeg

Gatekeeper blocks the app? Right-click MacWispr.app → Open, or:

xattr -dr com.apple.quarantine /Applications/MacWispr.app

App quits / bench fails with Failed to load the default metallib?
Two common causes:

Only Command Line Tools installed (no full Xcode.app) — swift build still succeeds, but MLX cannot load Metal shaders. Confirm:

xcrun -sdk macosx metal --version
# error: unable to find utility "metal"  → you need full Xcode

Fix:

# Install Xcode from the Mac App Store, then:
sudo xcode-select -s /Applications/Xcode.app/Contents/Developer
xcodebuild -runFirstLaunch
# if metal is still missing:
xcodebuild -downloadComponent MetalToolchain
./scripts/preflight-xcode.sh   # should succeed

Packaged .app missing mlx.metallib — rebuild with the packaging script (requires the Metal toolchain above):
```
./scripts/install.sh
```
scripts/build-app.sh compiles the metal kernels and places mlx.metallib next to the binary inside MacWispr.app.

License

MIT — see LICENSE.

Formerly OpenWhispr — renamed to MacWispr to better reflect the macOS-first product.

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
BenchLatency		BenchLatency
GTM		GTM
Sources		Sources
Tests/MacWisprCoreTests		Tests/MacWisprCoreTests
bench		bench
docs		docs
leaderboard-api		leaderboard-api
scripts		scripts
website-macwispr		website-macwispr
website		website
.gitignore		.gitignore
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
Info.plist		Info.plist
LICENSE		LICENSE
MacWispr.entitlements		MacWispr.entitlements
PRIVACY.md		PRIVACY.md
Package.resolved		Package.resolved
Package.swift		Package.swift
README.md		README.md
bench.sh		bench.sh

Navigation Menu

Repository files navigation

MacWispr

Website

Easy install (stable)

Beta (1.2.4-beta.1) — bundled on-device polish

One-command build & install (from source)

Features

Weekly Time Saved dashboard

Requirements

Usage

Benchmark

At a glance (Apple M5, 32 GB)

Latency detail

Accuracy detail (WER)

How to read this

Qwen3 family only (in-app size ladder)

Optimizations in MacWispr

Architecture

Dependencies

Troubleshooting

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

MacWispr

Website

Easy install (stable)

Beta (1.2.4-beta.1) — bundled on-device polish

One-command build & install (from source)

Features

Weekly Time Saved dashboard

Requirements

Usage

Benchmark

At a glance (Apple M5, 32 GB)

Latency detail

Accuracy detail (WER)

How to read this

Qwen3 family only (in-app size ladder)

Optimizations in MacWispr

Architecture

Dependencies

Troubleshooting

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages