A desktop video editor for spoken-word videos (talks, podcasts, tutorials, lectures and vlogs), built in Python on FFmpeg. It focuses on the edits that eat the most time: cutting dead air, fixing the soundtrack and pulling out the audio, all with a live waveform and exports that keep audio and video in sync.
It is built one feature at a time. The roadmap and the issue tracker list what's coming next, from playback preview to automatic subtitles.
Removing 226 silent parts (red) from a 21-minute recording: 41% of the length is cut.
About the name. The project started as a silence remover, and the code still uses that name: the Python package and the command are
silence_remover, and the window title is "Silence Remover". The commands below use these current names.
| Tool | What it does | GUI | CLI |
|---|---|---|---|
| Silence removal | Cuts pauses and dead air automatically, by loudness or by AI voice detection. Frame-accurate or lossless export. | Export… | INPUT OUTPUT [options] |
| Extract audio | Saves a video's full soundtrack as an MP3. | Export audio (MP3)… | --extract-audio |
| Swap audio | Replaces a video's soundtrack with another file, such as a cleaned-up recording. The picture is copied untouched. | Swap audio… | --swap-audio AUDIO |
Across all tools:
- Interactive waveform: zoom, pan and drag thresholds; cuts update live.
- Instant preview: each file is analyzed once, so every setting change shows immediately.
- Background processing: a progress bar and Cancel on every export; the window never freezes.
- GUI and CLI: every tool also works from the command line, for scripting and batch jobs.
- Tested: 105 automated tests, run on Windows and Linux in CI.
These are planned and tracked as issues, in rough order:
| Phase | Highlights |
|---|---|
| 1. Quick wins | Presets, drag and drop, remembered settings, more audio formats (WAV/M4A/FLAC), shortcuts |
| 2. Preview and editing | Playback preview, manual cut editing, undo/redo, export cut lists to Premiere/Resolve/Final Cut |
| 3. Audio quality | Noise reduction, loudness normalization, fades at cuts, shortening pauses instead of cutting |
| 4. Speed and scale | GPU encoding, export quality settings, a batch queue |
| 5. AI | Filler-word removal ("um", "uh"), automatic subtitles, editing by deleting transcript text |
| 6. Distribution | A standalone Windows installer with FFmpeg bundled, macOS support |
See ROADMAP.md for the details of each feature.
FFmpeg does all decoding and encoding, and ffmpeg and ffprobe must be on
your PATH.
ffmpeg -version # should print "ffmpeg version 7.x" or newerOn Windows, winget install Gyan.FFmpeg works, as do the builds from
https://www.gyan.dev/ffmpeg/builds/.
git clone https://github.com/ifahimkhan/Video-Editor.git
cd Video-Editor
python -m venv .venv
.venv\Scripts\activate # Windows
# source .venv/bin/activate # macOS / Linux
pip install -r requirements.txt| Package | Needed for |
|---|---|
numpy |
audio analysis (required) |
PyQt6 |
the desktop app |
pyqtgraph |
the interactive waveform (without it, the app uses a simpler static timeline) |
onnxruntime |
AI voice detection (without it, only loudness detection is available) |
pytest, pytest-cov, ruff |
development: tests and lint |
python -m silence_remover # open the editor
python -m silence_remover talk.mp4 talk_cut.mp4 # or: remove silence from the command line
Click Open video… and pick a video or audio file (mp4, mkv, mov, avi, webm, m4v, mp3, wav or m4a). The editor analyzes it once (a one-hour file takes a few seconds) and draws its waveform. Every tool then works on that file, and each export shows its progress and can be cancelled.
- Tune the controls. The waveform shows what will be kept and what will be
cut, shaded red. The line below it sums it up, e.g.
Original 12:04 → 9:31 removed 2:33 (21%) in 87 cuts. - Pick an export mode and click Export….
| Control | What it does | Tips |
|---|---|---|
| Detection | Loudness cuts anything quieter than the threshold. Voice keeps only what the Silero VAD model recognizes as speech. | Use Voice when there's background noise or music. |
| Threshold (dB) | Loudness mode: audio quieter than this counts as silence. Range −70 to −10, default −35. | Quiet room: −40 to −35. Noisy room: −30 to −25. |
| Voice sensitivity (%) | Voice mode: how confident the model must be that audio is speech. Range 5–95, default 50. | Lower keeps more (soft or whispered speech); higher cuts more aggressively. |
| Minimum silence (ms) | Only pauses at least this long are cut. Range 50–5000, default 500. | 300–500 for tight edits; 800+ to keep a natural pace. |
| Softness / padding (ms) | Audio kept before and after each stretch of speech so words aren't clipped. Range 0–1000, default 150. | Raise it if word beginnings or endings sound chopped. |
| Export | Precise or Fast lossless; see Export modes. |
- Light blue shows the peak level. Dark blue shows the RMS loudness, which is what the threshold is compared against. The waveform uses a dB scale, so quiet passages stay visible.
- Red shading marks the parts that will be removed.
- Orange dashed line: the threshold. Drag it and the slider follows.
- Yellow lane (Voice mode only): speech probability over time, with its own draggable sensitivity line.
- Mouse wheel zooms the time axis and dragging pans it.
| Loudness | Voice (Silero VAD) | |
|---|---|---|
| Decides by | RMS level in dBFS | Neural speech probability |
| Background hum, music, fans | Kept if louder than the threshold | Removed |
| Very quiet or whispered speech | Cut if below the threshold | Usually kept |
| Extra analysis time | none | about 2 s per 15 min of audio |
Voice mode uses the bundled model silence_remover/models/silero_vad_v6.onnx.
To use a different copy, point the SILENCE_REMOVER_VAD_MODEL environment
variable at it. Only this exact Silero v6 export (the one shipped with
faster-whisper) is supported.
| Precise (default) | Fast lossless | |
|---|---|---|
| How | One FFmpeg pass that re-encodes (H.264 CRF 18 + AAC 192k) | Stream copy, no re-encoding |
| Speed | Roughly as fast as your CPU encodes | Seconds, even for long files |
| Quality | Visually lossless re-encode | Bit-identical to the source |
| Cut accuracy | ~5 ms | Each kept part starts on the previous keyframe |
| Output format | MP4 or MKV | Same container as the source by default |
Why lossless keeps a little extra. Without re-encoding, video can only start
playing from a keyframe. Each kept part therefore begins at the keyframe just
before the speech: speech is never clipped, but some silence may remain. The
timeline previews exactly what lossless will keep, and the stats show the cost,
e.g. (+1.2s kept to start on keyframes).
Phone and camera footage usually has a keyframe every 0.5–2 s, which works well. Screen recordings often have only one every 10 s or more, so use Precise for those.
Export audio (MP3)… saves the video's complete original soundtrack as an MP3 (192 kbps, with title/artist tags copied).
- It is a plain extraction: nothing is cut, and the silence settings have no effect.
- The button becomes available once a file has been analyzed.
- From the command line you can also pick the bitrate (64–320 kbps).
Swap audio… replaces the video's soundtrack with another audio file, for example a version cleaned up in a noise-removal tool. Pick the new audio (wav, mp3, m4a, aac, flac, ogg or opus), then where to save.
- The video stream is copied untouched (bit-identical, no quality loss); only
the new audio is encoded (AAC 192 kbps, or Opus for
.webm). - The result always has the video's length. Shorter audio is padded with silence and longer audio is trimmed, with a warning when the two differ by more than half a second.
- No silence is removed. To also cut pauses, open the swapped video and remove silence as usual.
- If the new audio is slightly out of sync, the command line's
--audio-offsetshifts it. A GUI control is planned (#4).
python -m silence_remover # no arguments: open the editor
python -m silence_remover INPUT OUTPUT [options] # remove silence, extract or swap audio
python -m silence_remover -h # list all optionsWithout --extract-audio or --swap-audio, the command removes silence.
| Argument | Description |
|---|---|
INPUT |
Source video or audio file (mp4, mkv, mov, avi, webm, m4v, mp3, wav, m4a, …). |
OUTPUT |
File to write. The extension picks the container: .mp4/.mkv for video (use the source's extension with --swap-audio), .mp3 for --extract-audio. Required even with --dry-run, which never writes it. |
| Option | Default | Allowed values | Description |
|---|---|---|---|
-h, --help |
Show the help text and exit. | ||
| Silence removal | |||
--mode {loudness,voice} |
loudness |
loudness, voice |
Detection mode. voice uses Silero VAD and needs onnxruntime. |
-t, --threshold DB |
-35 |
−70 to −10 | Loudness mode: audio quieter than this (dBFS) is silence. |
-v, --vad-threshold P |
0.5 |
0.05 to 0.95 | Voice mode: speech probability above which audio counts as speech. |
-m, --min-silence MS |
500 |
50 to 5000 | Only pauses at least this long (ms) are cut. |
-p, --padding MS |
150 |
0 to 1000 | Softness: audio kept (ms) before and after speech. |
--lossless |
off | Fast export without re-encoding; each kept part starts on a keyframe. | |
--dry-run |
off | Analyze and print what would be cut; write nothing. | |
| Extract audio | |||
--extract-audio |
off | Save INPUT's full audio track as MP3 to OUTPUT. No silence removal, so the silence-removal options are ignored. | |
--mp3-bitrate KBPS |
192 |
64, 96, 128, 160, 192, 256, 320 | MP3 bitrate for --extract-audio. |
| Swap audio | |||
--swap-audio AUDIO |
path to an audio file | Replace INPUT's soundtrack with AUDIO and save to OUTPUT. The video is copied untouched and no silence is removed, so the silence-removal options are ignored. | |
--audio-offset MS |
0 |
−60000 to 60000 | With --swap-audio: shift the new audio later (positive) or earlier (negative). |
-t only applies in loudness mode and -v only in voice mode; the other
silence-removal options apply to both. --extract-audio and --swap-audio
are separate jobs and can't be combined.
# Silence removal with the defaults: loudness mode, -35 dB, 500 ms, 150 ms padding
python -m silence_remover talk.mp4 talk_cut.mp4
# Report what would be cut, without writing anything
python -m silence_remover lecture.mp4 out.mp4 --dry-run
# Quiet room: stricter threshold, keep more around words
python -m silence_remover podcast.mp4 podcast_cut.mp4 -t -40 -p 250
# Noisy recording: detect speech instead of loudness
python -m silence_remover vlog.mp4 vlog_cut.mp4 --mode voice -v 0.4
# Fast, lossless, with tighter pauses
python -m silence_remover interview.mov interview_cut.mov --lossless -m 300
# Everything at once
python -m silence_remover in.mp4 out.mp4 --mode voice -v 0.6 -m 400 -p 200 --lossless
# Extract the soundtrack as MP3 (default 192 kbps, or choose one)
python -m silence_remover lecture.mp4 lecture.mp3 --extract-audio
python -m silence_remover lecture.mp4 lecture.mp3 --extract-audio --mp3-bitrate 320
# Put the cleaned-up audio back on the video
python -m silence_remover talk.mp4 talk_clean.mp4 --swap-audio talk_clean.wav
# ...and shift it 120 ms later if it runs ahead of the picture
python -m silence_remover talk.mp4 talk_clean.mp4 --swap-audio talk_clean.wav --audio-offset 120
# Batch a folder (PowerShell)
Get-ChildItem *.mp4 | ForEach-Object { python -m silence_remover $_ "cut_$($_.Name)" --lossless }The exit code is 0 on success and 1 on errors. Errors include invalid
settings, a missing file, a file without the needed audio or video track,
everything being below the threshold, and voice mode without onnxruntime.
An unknown option, an out-of-list --mp3-bitrate, or combining
--extract-audio with --swap-audio makes argparse print usage and exit
with 2.
Every tool is a Qt-free module that builds FFmpeg commands; the GUI and the CLI are thin layers on top. That keeps each tool unit-testable and makes new features easy to add one at a time.
input ──ffmpeg──▶ 16 kHz mono PCM stream
├─▶ 10 ms RMS + peak levels (dBFS) ─┐
└─▶ Silero VAD, 32 ms speech probability ─┤ analyzed once
▼
settings ─▶ silent runs ≥ min silence ─▶ keep = speech + padding
│
┌───────────────────────────────────┴──────────────┐
Precise export Lossless export
one ffmpeg pass, select/aselect, snap starts to keyframes, copy
timestamps shifted by removed time, each segment in parallel,
filter graph loaded from a file join with concat demuxer
- Analysis streams audio out of FFmpeg in chunks, so memory use stays low even on multi-hour files. The Silero model state is carried from chunk to chunk, so results match running the model on the whole file at once.
- Detection is a set of pure functions. It finds silent stretches, drops those shorter than the minimum, then pads the remaining speech and merges any overlaps. Voice mode first switches between speech and non-speech with hysteresis (Silero's recommended 0.15 gap, so words aren't split) and ignores speech shorter than 100 ms.
- Precise export shifts every frame's timestamp back by exactly the amount
of silence removed before it. This keeps audio and video in sync, including
variable-frame-rate phone video.
- The filter graph goes in a file, so hundreds of cuts never hit the Windows command-line length limit.
- Long sums are nested in balanced pairs, because FFmpeg's expression parser rejects a flat chain of more than 100 terms.
- Lossless export copies each segment separately rather than cutting inside a single concat script. The single-script approach trims by decode timestamp, which leaks B-frames from the next segment and puts frames out of order at every join.
- Extract audio is one FFmpeg pass: first audio track →
libmp3lame, with the source's tags copied. - Swap audio maps the video stream from the original file (stream copy) and
the audio from the new file. It pads the audio with
apadand caps the output at the video's duration, so the result always matches the picture. An offset shifts the audio withadelay(later) oratrim(earlier).
silence_remover/
__main__.py entry point: editor without arguments, CLI with them
cli.py command-line interface
ffmpeg_tools.py ffmpeg/ffprobe discovery, version check, media probing
analysis.py streaming decode → loudness, peak and speech-probability arrays
vad.py streaming Silero VAD wrapper (onnxruntime)
segments.py silence detection (pure functions)
render.py precise export (re-encode) with progress and cancel
lossless.py keyframe scan, snapping, stream-copy export
audio_export.py extract audio: full audio track to MP3
swap_audio.py swap audio: replace the soundtrack, video stream-copied
models/
silero_vad_v6.onnx
gui/
main_window.py window, tool buttons, controls, live preview
waveform_timeline.py pyqtgraph waveform with draggable thresholds
waveform_data.py plot math (numpy only, unit tested)
timeline.py fallback timeline when pyqtgraph is missing
workers.py QThread workers for analysis and every export
docs/ screenshots used in this README
tests/
fixtures/speech.flac short synthetic speech clip for VAD tests
ROADMAP.md planned features and how we add them
New features are added one at a time, each on its own branch with tests, a
pull request and green CI. ROADMAP.md describes the workflow and
every planned feature, and each one has an
issue labelled
enhancement, with its phase and effort. The good first issue label marks
small starting points.
python -m pytest -q # 105 tests
python -m pytest -q --cov=silence_remover # with coverage (~88%)
ruff check . # lint (config in ruff.toml)- The integration tests generate their own clips with FFmpeg and are skipped if FFmpeg is missing.
- CI (
.github/workflows/ci.yml) runs lint plus the tests on Ubuntu (Python 3.10, 3.12, 3.13) and Windows (3.12), with a static FFmpeg 9.0 build and an 80% coverage gate. - The GUI tests run offscreen (
QT_QPA_PLATFORM=offscreen), so no display is needed. - The tests cover, among others:
- A/V sync after 60 cuts, and precise export with more than 100 cuts
- Voice mode removing loud noise that Loudness mode keeps
- Lossless output being bit-identical to the source and decoding without errors at every join
- Swapped audio really replacing the soundtrack, with the picture untouched
- The full threaded GUI flow for every tool
| Problem | Fix |
|---|---|
'ffmpeg' not found on PATH |
Install FFmpeg and reopen the terminal. |
FFmpeg 6 is too old |
Export needs FFmpeg 7+ (it uses -/filter_complex <file>). |
The selected file has no audio track |
The editor needs an audio track to analyze a file. For a silent video, only Swap audio can add a soundtrack, and that currently runs from the command line only. |
| Voice option is greyed out | Install onnxruntime and check that silence_remover/models/silero_vad_v6.onnx exists. |
| Word beginnings or endings are clipped | Raise Softness / padding, or lower the threshold / voice sensitivity. |
| Background noise isn't removed | Switch to Voice detection. |
| Lossless keeps too much silence | The source has few keyframes (common for screen recordings); use Precise. |
| Lossless export fails with a muxer or codec error | The source codecs don't fit the chosen container; save with the source's own extension or .mkv. |
| Swapped audio is out of sync | Use --audio-offset (positive = later) on the command line. |
pip fails with CERTIFICATE_VERIFY_FAILED ... self-signed certificate |
Antivirus or a proxy is inspecting HTTPS. Export the Windows trusted roots into a PEM bundle and run pip install --cert bundle.pem -r requirements.txt. |
- FFmpeg for decoding, filtering and encoding.
- Silero VAD (MIT license). The bundled ONNX file is the v6 export distributed with faster-whisper (MIT license).
- PyQt6 and pyqtgraph for the GUI.