1
Drop in a recording
WAV, MP3, M4A, FLAC, MOV, MP4 and other common formats. Each project is one
.vc document you open from the Finder, your file manager or Explorer.
Free · Open source · macOS, Windows, Linux
VoiceCut
Record with all your retakes.
Get back one clean take.
VoiceCut listens to your narration, keeps the best take of every line, drops false starts, repeats and breaths, and gives every pause a natural length. Then you fine-tune every cut on a timeline and export audio or video.
What it does
It hears the retakes and keeps the take you meant
The video from the screenshots, word for word. Struck-through words are what Voice Cut removed; nothing was added or regenerated.
Removed: We can do.we can rewrite the same loss in the row-wise manner. Every element of the sum here is an L2 loss of the corresponding row from the matrix. That means we can study one row at a time.Removed: thatThis makes the problem much simpler. since if we can solve this problem for a single row,Removed: then we can solve this forThen we can solve this for all rows just by applying the same algorithm to every other row.
-
The best take of every line
It tells a corrected retake from a fluent but abandoned attempt, and keeps the one you meant.
-
Cuts only where it is silent
Every cut lands between words, never inside one, and keeps each word's natural decay.
-
Natural pauses
Each pause gets a length for its kind — set once in Settings for every project — and breaths and clicks become clean room tone.
-
Your real voice
Nothing is synthesized. Every sound in the edit comes from your own recording.
-
Video too
The picture follows the kept speech with direct cuts — no frozen frames, no fake pauses.
-
Runs on your computer
Transcription, alignment and rendering are local, and the planner can be too: Gemma 4 on your own Mac or PC, the fast Kev on a Mac with less memory, or a cloud model with your own key.
2
Press Voice Cut
Pick the planner, its model and the language, then press one button. Every stage shows its progress and the run the time left, learned from your own computer's speed. An interrupted run picks up where it stopped, and trying another model runs only the take selection again.
3
Fine-tune every cut
The edit opens as clips over the original recording, so nothing is ever lost.
- Drag a clip edge to bring back removed words or trim them
- Resize pauses, or press P to insert one
- Set how long each kind of pause is, once for every project
- Drag clips to rearrange them, or copy and paste them
- Split with S, delete, undo and redo
- Listen at 0.5× to 2.5× while the caption follows every word
4
Export what you hear
MP4 or MOV video, or WAV, FLAC, AAC and MP3 audio. The preview is rendered by the same code as the export, so the file sounds exactly like the editor.
Real examples
Hear the difference
Straight out of Voice Cut, with no manual repair. The bars are drawn to scale: the shorter one is what is left once the retakes are gone.
English narration
False starts become one clean take
Original0:38
VoiceCut0:27
The abandoned opening and the repeated phrase are gone; the complete take stays.
Русская речь
Повторы превращаются в связное объяснение
Оригинал0:35
VoiceCut0:22
Неудачные попытки и повторённые фразы удалены, финальное объяснение сохранено.
Video
The picture follows the kept speech
Direct cuts at the kept speech: no frozen frames and no artificial pauses.
Planners
Choose the model that edits
The planner reads the transcript and chooses the takes. Its default model, on your computer and on OpenRouter, is Gemma 4 31B: an open model that edits nearly the way Gemini 3.1 Pro does, so a big model is not needed. New projects plan on your computer where it runs Gemma 4 well, with the faster Kev decision models on a Mac with less memory, and through OpenRouter elsewhere. Other providers work too, and you can type any model name.
-
Local (Ollama)On your computer
Gemma 4 31B by default, on your Mac or PC, downloaded in the app. Nothing leaves the computer; it needs about 32 GB of memory.
-
Kev 9B and Kev 4BDecision models
Fast and light, on a Mac with Apple Silicon from 12 GB of memory: an hour of narration plans in about 10–20 minutes, a little less precisely than Gemma 4. They answer typed questions about every line instead of writing text. TypeSafe Jev, a hosted one, is in the OpenRouter list.
-
OpenRouterGemma 4
The same Gemma 4 31B by default, in the cloud. It reads the transcript of every window.
-
GeminiHears the audio
Gemini 3.8 Flash by default. It hears the audio of every window as well as reading it, so it catches stumbles and mis-heard words.
-
OpenAISupported
Reads the transcript of every window.
-
DeepSeekSupported
Reads the transcript of every window.
VoiceCut is free. A cloud planner bills your own API key at its provider, and none of it goes to VoiceCut: for an hour of recording, about 5 cents with Gemma 4 through OpenRouter, about 3 cents with Jev and about $1 with Gemini 3.8 Flash, at their list prices. Your original file and any video never leave the computer. Keys live in the macOS Keychain, the Windows Credential Manager or your Linux keyring, and are sent only to their own provider.
Tested models
How the models compare
Each model chose the takes of the same 64-minute narration, and its edit was compared with the author's own. A score of 1.00 is the author's cut; cutting everything scores 0.36.
| Model | Score | Per 10 min of recording |
|---|---|---|
| Gemini 3.8 FlashHears the audio Gemini API | 0.974 | 3 min |
| Gemini 3.8 Flash Gemini API, transcript only | 0.964 | 4 min |
| Gemini 3.1 Pro OpenRouter | 0.955 | 4.4 min |
| Qwen 3.8 27BThinking OpenRouter | 0.950 | 25 min |
| GLM-5.3 OpenRouter, low thinking | 0.923 | 1.1 min |
| Gemma 4 31BDefault OpenRouter | 0.910 | 1.3 min |
| Gemma 4 31B, 4-bitDefault Your computer (Ollama): A100 40 GB or M1 Max 64 GB | 0.903 | 2.8 min on the A100, 16 min on the M1 Max |
| TypeSafe Jev 1.14Decision OpenRouter | 0.889 | 4 s |
| Qwen 3.8 27B, 4-bit Your computer (Ollama): A100 40 GB | 0.887 | 10 min |
| Kev 9BDecision Your Mac (Apple Silicon): M1 Max 64 GB | 0.854 | 3 min |
| Qwen 3.6 35B A3BThinking OpenRouter | 0.852 | 4.8 min |
| Kev 4BDecision Your Mac (Apple Silicon): M1 Max 64 GB | 0.818 | 1.5 min |
| Qwen 3.5 122B A10B OpenRouter | 0.773 | 1 min |
Decision models were scored on two 10-minute parts of the recording, an easy one and a hard one, where Gemma 4 31B scored 0.90–0.93. Times are for choosing the takes only; in the cloud they depend on how busy the provider is. Most models of 14B parameters or less could not produce a valid edit at all. Every model and how they were measured are in the planner benchmark.
How it works
Choose the takes, then cut only where it is silent
-
1
Transcribe everything
Whisper writes down every word, and speech it skipped is transcribed again on its own.
-
2
Choose the takes
The planner reads minutes at a time with look-ahead, and revokes a choice when a better retake comes.
-
3
Find the real edges
The Montreal Forced Aligner and the waveform decide where every kept word starts and stops.
-
4
Fix the pauses
Every gap gets a natural length; breaths and clicks become room tone from the same recording.
-
5
Render once
Kept speech is copied sample for sample and joined with short crossfades in silence.
macOS
macOS 14 Sonoma or later
Apple Silicon or Intel
Windows
Windows 10 or 11
x64
Linux
x86_64 with glibc 2.35+
Ubuntu 22.04, Debian 12, Fedora 36 or newer
Opening it the first time
-
Allow the download
The installers are not signed with paid certificates yet, so your system asks once. macOS: System Settings › Privacy & Security › Open Anyway. Windows: More info › Run anyway. Linux: make the AppImage executable; if it does not start, install
libfuse2. -
Install the engine
One click sets up Python, PyTorch, Whisper, the aligner and FFmpeg: about 1.5 GB, once. Speech models, about 2 GB, follow with the first Voice Cut.
-
Choose a planner
On a Mac with Apple Silicon and 32 GB, or a PC with a 24 GB NVIDIA GPU, VoiceCut plans on your computer and downloads Gemma 4 with the first Voice Cut. Elsewhere paste an OpenRouter key in Settings for the same model in the cloud, or choose Gemini, OpenAI or DeepSeek.
Questions
Is VoiceCut really free?
Yes. VoiceCut is open source under the MIT License, with no account and no subscription, and with a planner on your computer it costs nothing at all. A cloud planner uses your own API key, so its provider bills you for what Voice Cut sends it — none of it goes to VoiceCut. For an hour of recording that is about 5 cents with Gemma 4 through OpenRouter.
macOS says VoiceCut can't be opened. What do I do?
The installers are not signed with a paid Apple certificate yet, so macOS stops the first launch of
a downloaded copy. Drag VoiceCut to Applications first (a copy opened from the disk image cannot
update itself), open it, and close the message. Then open System Settings › Privacy & Security,
scroll down to the note that VoiceCut was blocked, choose Open Anyway and confirm. You do
this once. In Terminal, xattr -dr com.apple.quarantine /Applications/VoiceCut.app does
the same.
And on Windows or Linux?
On Windows, SmartScreen says it protected your PC: choose More info › Run anyway.
The installer needs no administrator. On Linux, make the AppImage executable
(chmod +x) and run it; if it does not start on Ubuntu, install libfuse2
(libfuse2t64 on 24.04 and later). The .deb installs with
sudo apt install ./VoiceCut-…-amd64.deb.
Does my recording leave my computer?
Not with a planner on your computer: then everything runs locally. With a cloud planner, transcription, alignment and rendering still run locally; the planner receives the transcript of each planning window, and Gemini also a compressed 16 kHz mono copy of that window's audio. Your original file and any video never leave the computer.
Which languages does it understand?
English and Russian, recorded by one main narrator. You choose the language in the toolbar before pressing Voice Cut.
Will it ever cut a word in half?
Cuts are placed where the waveform is silent, between words. Tightly connected speech, with no pause between a kept and a removed word, can occasionally sound abrupt, and the planner can occasionally choose the wrong take. Every cut can be adjusted on the timeline, so listen before you publish.
What does it need to run?
macOS 14 or later (Apple Silicon or Intel), Windows 10 or 11 on x64, or Linux on x86_64 with glibc 2.35+, and about 4 GB of disk for the engine. Whisper uses the GPU on Apple Silicon and on NVIDIA GPUs; elsewhere it runs on the CPU, about as fast as the recording plays.
How do updates work?
When a new version is out, a download button appears in the toolbar. By default the update downloads in the background and installs when you quit, and your projects open again after the restart. You can switch that off in Settings.
Is there a command line?
Yes: voicecut recording.wav runs the same pipeline from a source checkout. See
the development guide.