sayszz

What does sayszz do?

sayszz is a Python-packaged dictation tool. Tap the Fn key (or a hotkey you pick), speak, and the words appear where the cursor is in any app. The README states: "No cloud, no subscription, no audio ever leaving your machine."

The feature set lists one-key dictation (tap or push-to-talk), paste at the cursor in whatever app is focused, Metal acceleration on Apple Silicon through MLX with a CPU fallback through faster-whisper, a searchable local history web UI with total words, live words-per-minute and day-streak stats, a menu-bar control, voice commands for keystrokes, snippets and shell commands, smart silence detection via TEN VAD, and an interactive sayszz-setup wizard that handles the hotkey, model, icon colour, Dock pin and start-at-login.

Key facts

• Licence: MIT on the code.
• Platforms: macOS 15 or later (preferred) and Windows (CPU path). Apple Silicon supports MLX Metal acceleration; Intel Macs and Windows run on faster-whisper CPU int8.
• Transcription engines: MLX (mlx_engine.py) on Apple Silicon via Metal GPU, and faster-whisper (whisper_engine.py) on CPU and CUDA.
• Models: Whisper in sizes tiny, base, small, medium and large-v3-turbo. First run downloads the chosen Whisper model (75 MB for tiny).
• History: a local SQLite log and a web UI on http://127.0.0.1:8765, which binds to localhost and is not reachable from the network.
• Settings: ~/.sayszz/user_settings.yaml for hotkeys, engine, models, VAD thresholds, paste behaviour and audio feedback.
• Requires Python 3.11 to 3.13.
• Credits whisper-key-local by Pin Wang (MIT) as the cross-platform hotkey, recording and paste foundation.
• The only outbound network call is a Hugging Face model download on first use.

What does it replace, and where does it fall short?

sayszz replaces Wispr Flow and Superwhisper for the one-key dictation loop with everything on the machine. The README names Wispr Flow directly as the cloud dictation tool it mirrors.

Where it falls short: Windows support is CPU only, so a slower setup. Install is through Python packaging (pip install -e .), not a DMG or installer, so a user has to be comfortable with a virtual environment. There is no AI cleanup step; the output is what Whisper returns. The history file lives in ~/.sayszz/history.jsonl and settings in ~/.sayszz/user_settings.yaml, both unencrypted on disk. The project is a solo build.

How does sayszz run?

Clone the repository, create a virtual environment, install with pip install -e ., run sayszz-setup for the wizard, then run sayszz to start dictating. First run on macOS asks for Microphone and Accessibility permissions. The README suggests freeing the Fn key from the emoji picker: System Settings, Keyboard, Press 🌐 key to Do Nothing.

The engine picks itself: MLX on Apple Silicon, faster-whisper on Intel macOS or Windows. The hotkey and recording mode are set in the YAML config file. Modes are toggle or push_to_talk. A separate sayszz-history process serves the local web UI; the setup wizard can pin it to the Dock with a chosen icon colour.

Who is sayszz for?

Someone on an Apple Silicon Mac who wants dictation with MLX Metal acceleration and a readable history, with no account and no cloud. A Windows user with a CPU can still run it, more slowly. Someone who wants a signed installer and a shipped macOS app bundle without a Python environment should pick Handy or BetterWispr on the open-source alternatives to Wispr Flow list.

What limits does the README state?

Requires Python 3.11 to 3.13. Windows speech-to-text runs on CPU through faster-whisper. The audio is processed in memory and transcripts are stored only in ~/.sayszz/history.jsonl. The history UI binds to 127.0.0.1 and is not reachable from the network. The only kind of outbound request is downloading models from Hugging Face on first use. Larger Whisper models trade speed for accuracy; tiny is the fastest and smallest.

Questions people ask

Is sayszz open source?

Yes. The code is MIT, an OSI approved licence. The README states: "free to use, fork, and ship." Credits are given to whisper-key-local by Pin Wang (MIT), on which sayszz builds, and to OpenAI's Whisper models, Apple's MLX and SYSTRAN's faster-whisper.

Can sayszz be used commercially?

Yes. MIT allows commercial use, modification and redistribution, provided the licence travels with the code. Transcription models are Whisper (MIT weights by OpenAI); the project ships no model of its own. Nothing is sent to a hosted provider by default.

Does sayszz work offline?

Yes, after the first model download. The README states that sayszz makes exactly one kind of network request: downloading models from Hugging Face on first use. Audio is processed in memory on the machine, and the history web UI binds to localhost only, so it is not reachable from the local network.

How does sayszz compare with Wispr Flow?

Wispr Flow is a hosted product with a monthly subscription, where speech is sent to the vendor's servers. sayszz runs the Whisper model on the machine with MLX on Apple Silicon or faster-whisper everywhere else, and nothing leaves the computer after the first download. The open-source alternatives to Wispr Flow page sets it beside Handy, VoiceInk and OpenWhispr.

How accurate is sayszz, and which model should I pick?

Accuracy is Whisper's own, not sayszz's. The README's model list runs from tiny (fastest, smallest) to large-v3-turbo. Bigger models trade speed for accuracy and stay interactive on M-series chips. The maintainer's own figure, on an Apple M5 MacBook Air transcribing 11.3 seconds of speech with the tiny model, is 0.13 seconds on MLX and 0.83 seconds on faster-whisper (CPU int8). Those are the project's own numbers.

Sources

• sayszz README and LICENSE: github.com/piyushxt43/sayszz, read 2026-10-26. The README is the source for every feature, model, benchmark figure and limit above.