Howl

Howl

Open-source voice dictation for Mac. Hold a hotkey, speak, release — your words type themselves. Whisper transcribes locally, your LLM cleans up the rest.

Free, privacy-first alternative to Wispr Flow, Superwhisper, and Voibe.

Download for MacView on GitHub

macOS 14+ · Apple Silicon · MIT license

Compare

How Howl stacks up

 HowlWispr FlowSuperwhisperVoibe
PriceFree$15/mo$250 lifetime$99 lifetime
Runs offline✓—✓✓
Filler-word removal✓✓——
Bring your own LLM✓———
Open source✓———
Custom dictionary✓partial——

Pipeline

How it works

Every stage is a swappable component selected by a preset. Bundled presets ship in the app; save your own from inside the editor or via the howl-cli tool.

  1. 01
    Mic
    Capture at 48 kHz
  2. 02
    VAD
    Detect speech (Silero)
  3. 03
    Denoise
    DeepFilterNet2 or RNNoise
  4. 04
    Whisper
    Local transcription (whisper.cpp)
  5. 05
    Dictionary
    Fix niche jargon
  6. 06
    LLM
    Clean fillers + grammar
  7. 07
    Text
    Inject at cursor

Features

What's in the box

Local Whisper transcription

Audio never leaves your machine unless you point Howl at a cloud LLM for cleanup. Picks the Whisper model size that fits your hardware.

Bring your own LLM

Anthropic Claude, OpenAI GPT, Ollama, or LM Studio. Use a $0 local model, or a state-of-the-art API. Your call, per preset.

Filler-word cleanup

Turn “um, so, like, basically what I mean is…” into the sentence you actually meant. Configurable per preset.

Custom dictionary

Fuzzy + phonetic matching preserves niche jargon — acronyms, brand names, code identifiers — that the base Whisper model gets wrong.

Hold-to-talk hotkey

Press, speak, release. Text injects at the cursor in any Mac app via the Accessibility API. Or toggle always-on mode.

Pluggable pipeline

VAD, noise suppression, ASR, dictionary, LLM cleanup are all swappable stages. A/B compare presets against the same captured audio.

LLM

Pick your model

Whisper transcription always runs locally. Cleanup goes through the provider you configure — cloud or local, your call.

Anthropic Claude
Best cleanup quality. Brings your API key.
cloud
OpenAI GPT
Familiar models, fast. Brings your API key.
cloud
Ollama
Fully offline. Local LLM, your hardware.
local
LM Studio
Fully offline. GUI for local model management.
local

FAQ

Common questions

Is Howl really free?+

Yes. MIT-licensed, no paid tier, no telemetry. The only thing you pay for is your own LLM API usage if you choose a cloud provider — and that’s optional. Run it 100% offline with Ollama or LM Studio for $0.

How is Howl different from Wispr Flow?+

Wispr Flow is a $15/month cloud service. Your audio leaves your machine to be transcribed and cleaned up on their servers. Howl runs Whisper locally and lets you decide whether the cleanup pass goes to a local LLM (zero data leaves your machine) or a cloud LLM you control (you bring the key).

How is Howl different from Superwhisper or Voibe?+

Both are paid, closed-source, and don’t remove filler words. Howl is free, open source, removes fillers automatically, and supports a custom dictionary for niche vocabulary.

Which platforms are supported?+

macOS 14+ on Apple Silicon today. The Go core is platform-agnostic; Linux and Windows wrappers are on the roadmap, not started.

Why “Howl”?+

Short, memorable, voice-native. Huskies howl. So does this thing — just into your keyboard.