Filler-Cut

A hardware-agnostic (AMD / Intel / NVIDIA) CLI tool that detects and cuts filler words ("um", "uh" — Turkish: "ııı", "şey", "yani") and unnecessary silences from video files using speech analysis.

v0.1 — see DESIGN.md for the architecture. Türkçe: README.tr.md

Windows application (installer)

Download: grab Filler-Cut-Setup-<version>.exe from the Releases page.

The build is unsigned, so Windows SmartScreen (and Smart App Control, if you have it on) may block the first run with "Windows protected your PC". This warning is expected (it flags the missing code signature, not malware): choose More info → Run anyway. If Smart App Control is enabled it can refuse outright — turning it off is a system-wide decision, so the alternative is running from source (see Install).

The lack of a signature is an accepted trade-off; an application to SignPath (free code signing for open-source projects) is under review. Once signed, this warning goes away.

Filler-Cut-Setup-<version>.exe installs per-user (no admin, no UAC) into %LOCALAPPDATA%\Programs\Filler-Cut and adds a Start Menu entry that opens the interface directly. The installer speaks Turkish and English, and resolves both prerequisites:

• WebView2 — runs Microsoft's official Evergreen Bootstrapper if the runtime is missing. If that fails the install still completes, with a warning that Filler-Cut will fall back to your browser.
• ffmpeg — not bundled (licence groups differ). If it is missing the finish page says so and offers winget install ffmpeg, or a manual link when winget is unavailable. It never blocks the install.

Uninstalling keeps your downloaded model. The program folder is removed, but %LOCALAPPDATA%\fillercut (whisper.cpp binary + model, ~570 MB) and your settings stay. The uninstaller asks whether to delete them — default no.

.\scripts\build_setup.ps1        # exe build + installer -> dist_setup\

The executables

scripts/build_exe.ps1 alone produces the standalone folder the installer ships:

exe	what
fillercut.exe	the console CLI — every command documented below
fillercut-ui.exe	no console; opens the interface directly

The packaged build defaults to the whisper.cpp (Vulkan) backend, so the first launch runs the setup wizard and then uses GPU acceleration on AMD, Intel and NVIDIA alike. pip install users are unaffected — there the default is still faster-whisper.

ffmpeg is not bundled; it stays a system dependency (see Requirements). The executables are unsigned, so SmartScreen may warn on first launch.

.\scripts\build_exe.ps1        # clean build + smoke tests -> dist\fillercut

Third-party components are listed in packaging/THIRD_PARTY_NOTICES.md, which the installer also copies next to the executables.

Requirements

• Python ≥ 3.10
• ffmpeg and ffprobe on PATH (system dependency — download)

Install

python -m venv .venv
# Windows: .venv\Scripts\activate
pip install -e .            # the CLI itself
pip install -e ".[cuda]"    # NVIDIA acceleration (cuBLAS/cuDNN for faster-whisper)
pip install -e ".[dev]"     # development: pytest, ruff, mypy

Backend & hardware support

Hardware	faster-whisper (default)	whispercpp
NVIDIA GPU	✅ CUDA (official wheel)	✅ official cublas package
CPU (everyone)	✅ int8	✅ official bin-x64 package
AMD GPU	❌ CTranslate2 has no ROCm support	✅ Filler-Cut Vulkan build (see below) or GGML_HIP=ON build (ROCm 7+)
Intel GPU	❌	✅ Filler-Cut Vulkan build (see below)

Note: upstream whisper.cpp Windows releases ship no Vulkan/HIP binaries (see issue #3673). Filler-Cut fills that gap with its own workflow: .github/workflows/vulkan-build.yml (whisper.cpp v1.9.1, -DGGML_VULKAN=ON) — on a v* tag push it builds and attaches the fillercut-whisper-cli-vulkan-win-x64.zip to the Releases page (permanent, no login needed); it can also be triggered manually from the Actions tab (artifact only). The package is vendor-agnostic: one binary for NVIDIA/AMD/Intel. On an RTX 4050 it matched CUDA speed (see KNOWN_ISSUES.md KI-1); only the very first run pays a one-time ~10 s shader compilation. No code changes are needed on the Filler-Cut side — the binary path comes from the whispercpp_binary config key.

Vulkan package setup (releases/v0.3.0+)

For AMD/Intel users who want GPU acceleration (or NVIDIA users who don't want a CUDA install), a ready-made package is on the Releases page:

1. Download fillercut-whisper-cli-vulkan-win-x64.zip, extract it (e.g. C:\tools\fillercut-whisper\).

2. Download the model: ggerganov/whisper.cpp → ggml-large-v3-turbo-q5_0.bin (~1.6 GB).

3. Create filler-cut.toml yourself in the folder where you run Filler-Cut (not shipped with the repo; paths are machine-specific, gitignored):

   config_version = 1
   
   [asr]
   backend = "whispercpp"
   whispercpp_binary = 'C:\tools\fillercut-whisper\whisper-cli.exe'
   whispercpp_model = 'C:\modeller\ggml-large-v3-turbo-q5_0.bin'
   

4. fillercut video.mp4 — that's it.

The first run pays a one-time ~10 s shader compilation (cached to disk). Proof the GPU is active is the ggml_vulkan: Found 1 Vulkan devices line in the output. No CUDA Toolkit / Vulkan SDK needed; an up-to-date GPU driver is enough.

Usage

fillercut video.mp4

Outputs, written next to the input (or to --output):

• video_temiz.mp4 — the cut video
• video_temiz.json — the cut report (every cut with its reason chain)
• video_transkript.json — the word-level transcript (kept even if you decline at the review step)

Options (identical to fillercut --help, which is Turkish — the CLI is Turkish-only; these are one-to-one translations):

--config PATH      TOML config file (default: filler-cut.toml).
--aggressive       Also cut candidate fillers (şey, yani, hani, işte).
-y, --yes          Skip the review confirmation (render without asking).
-o, --output PATH  Output MP4 path (default: <name>_temiz.mp4).
--open             Open the review HTML in the default browser once written.
--interactive      Approve cuts one by one in the browser (local server, v0.3).
--cikti mp4|xml    Output branch: rendered video (default) or NLE project.
--srt              Also write the transcript as <video name>.srt.
--version          Print the version and exit.

Before rendering, a review summary is printed and confirmation is asked (skipped with --yes) — real output from a 15 s test clip:

[1/6] EXTRACT — 16 kHz mono WAV çıkarılıyor…
[2/6] TRANSCRIBE — transkript çıkarılıyor…
[3/6] DETECT — filler ve sessizlikler tespit ediliyor…
[4/6] PLAN — kesim planı kuruluyor…
[5/6] REVIEW
Kesim sayısı: 4
Kademe dağılımı: 1 kesin filler, 0 aday filler, 4 sessizlik
Kazanılan süre: 00:03 (00:14 → 00:11), %22.28
                 İlk 5 kesim
┌───┬───────────┬───────┬─────────┬─────────────────────────────────────┐
│ # │ Başlangıç │ Bitiş │ Tür     │ Neden (reason)                      │
├───┼───────────┼───────┼─────────┼─────────────────────────────────────┤
│ 1 │ 00:03     │ 00:04 │ filler  │ sessizlik 1018ms (…) + kesin        │
│   │           │       │         │ filler: 'Eee,' [padding +80/-120ms] │
│ 2 │ 00:06     │ 00:07 │ silence │ sessizlik 704ms (…)                 │
└───┴───────────┴───────┴─────────┴─────────────────────────────────────┘
Render edilsin mi? [y/N]:
[6/6] RENDER — segmentler encode ediliyor…
Bitti: konusma_temiz.mp4 (%22.28 kazanım)
rapor: konusma_temiz.json
transkript: konusma_transkript.json

The first run downloads the Whisper model (~1 GB); later runs use the cache.

NLE project (FCP7 XML) and subtitles

If you would rather fine-tune the cuts in your own editor, --cikti xml renders nothing and writes video.xml instead:

fillercut video.mp4 --cikti xml --srt -y

• video.xml — an FCP7 (xmeml) timeline. Open it with File > Import in Premiere or DaVinci Resolve; it references the source video, so there is no re-encode and you can drag the cut boundaries around. RENDER never runs on this branch.
• video.srt — a standard subtitle file (--srt). Both land in the folder given by --output. Subtitles are on the cut timeline: words inside a removed region are dropped and the rest shift earlier, so the file drops straight onto the rendered video or the XML timeline. The source-timed record stays in video_transkript.json.

Cut boundaries are snapped to frames in favour of speech: a segment's start rounds down and its end rounds up. No syllable is ever clipped; at most one frame spills into the cut.

Example video_temiz.json (truncated):

{
  "original": { "ms": 14814, "human": "00:14" },
  "cut_total": { "ms": 3300, "human": "00:03" },
  "remaining": { "ms": 11514, "human": "00:11" },
  "saved_percent": 22.28,
  "cut_count": 4,
  "tiers": { "kesin_filler": 1, "aday_filler": 0, "silence": 4 },
  "cuts": [
    {
      "start_ms": 3164,
      "end_ms": 4182,
      "duration_ms": 1018,
      "kind": "filler",
      "reason": "sessizlik 1018ms (noise=-35dB, min=0.4s) + kesin filler: 'Eee,' [padding +80/-120ms]"
    }
  ]
}

Web UI

fillercut ui

On first launch, if the whisper.cpp engine or a model is missing, a setup wizard appears: pick a model, press one button, and it downloads (with a progress bar, resume on interruption and SHA-256 verification) into %LOCALAPPDATA%\fillercut. Jobs cannot start until it finishes. Prefer the terminal? fillercut setup does the same, and fillercut setup --durum reports what is installed and where it came from.

Opens Filler-Cut in its own desktop window (pywebview + the Windows WebView2 runtime), backed by a local server on http://127.0.0.1:8765 (loopback only — never binds beyond localhost). Without WebView2 — or without the optional pywebview package — nothing breaks: it falls back to your browser and prints one line saying why. Pick a video with the server-side file browser (files are not uploaded; the tool reads them from disk, and browsing is confined to your home directory), choose Normal/Aggressive mode, and watch the 6-stage pipeline progress live.

For the native window: pip install "fillercut[native]".

If your videos live on another drive (D:, E:\ …), the browser is confined to your home directory by default. Extend the jail by adding allowed roots in filler-cut.toml:

config_version = 1
[ui]
izinli_kokler = ['D:\', 'E:\Videolar']

If you'd rather not list drives one by one, izinli_kokler = ["*"] allows every mounted drive on the machine; the list is refreshed per request (not frozen at startup), so a USB disk plugged in later shows up on its own.

⚠️ Security: "*" lists all your drives to the localhost UI; it is not recommended on a shared or untrusted machine — list drives individually there.

These roots are read only from the config file — there is no UI control that changes them, so the security boundary can't be moved from inside the page. A nonexistent root makes fillercut ui stop with a clear error. With more than one root, a root switcher (Ev / D:\ …) appears above the browser.

You can also drag and drop a video: in the desktop window, drop the file onto the page or press "Dosya seç…" to open the native Windows file dialog. Drag and drop does not work in browser mode — browsers do not hand the page a file's disk path — and the UI says so and points you at the file browser, which is the picker in that mode.

After PLAN the run pauses for review: you get the video with a waveform timeline, every cut drawn on it, and a cut list. There you can

• play with skip mode on (cuts are skipped) or off (hear the original),
• undo any cut with one click — it stays in the list, greyed out, and one more click brings it back,
• drag a cut boundary; it snaps to the nearest silence edge,
• drag on empty timeline to add a cut of your own,
• snap a cut to silence with one click (Y) — the "Sessizliğe yasla" button on every row pushes both boundaries outward to the first silence edge, at most 500 ms per direction; it stops early at a neighbouring cut, so cuts never merge,
• turn the magnet off (M) when you want a boundary exactly where you drop it — snapping is on by default, and the toggle is shown in the header.

The header keeps a live line — how many cuts, how much will be removed, what the new duration will be — so you see the gain before you commit to it.

Approving renders. The result screen shows the output path, the time saved, a breakdown by cut type (definite/candidate filler, silence, your own cuts), and which filler words were cut (eee ×3, ııı ×1…) — plus a "show in folder" button for each output. Your edits are recorded in the report too (tiers.manuel, duzenleme).

Options: --port (default 8765), --config PATH (the same filler-cut.toml the CLI uses), --no-native (force browser mode), --native (require the native window — error out if unavailable), --no-browser (start the server, open nothing).

First-run setup

fillercut setup

Downloads the Vulkan whisper-cli build (from this repo's releases) and a GGML model (from ggerganov/whisper.cpp on Hugging Face). Options: --model NAME to choose a model, --yes for unattended/CI, --durum to report status instead of downloading.

model	size	when
ggml-large-v3-turbo-q5_0	547 MB	recommended — speed/accuracy balance
ggml-small-q5_1	190 MB	slow connection or tight disk
ggml-large-v3-q5_0	1.08 GB	quality-weighted, slowest

Paths resolve in this order, first existing candidate wins: filler-cut.toml → FILLERCUT_WCPP_BINARY/FILLERCUT_WCPP_MODEL → the wizard's own %APPDATA%\fillercut\config.json. So an existing setup never sees the wizard, and the wizard never overwrites your configuration.

The wizard installs the Vulkan build only — one binary for AMD, Intel and NVIDIA. The CUDA path stays manual for advanced users (see [asr].whispercpp_binary). ffmpeg remains a system dependency.

If port 8765 is busy the run does not fail: it falls back to a free port and tells you which one. If Filler-Cut is already running on that port, a second fillercut ui does not start a second server — it prints the address of the one that is already up.

Approving without any edits produces byte-for-byte the same file as the CLI run — the review screen adds control, not a different renderer.

Jobs live in memory only — restarting the server drops them (the tool tells you so instead of hanging). Rendered files stay on disk.

License

MIT — see LICENSE.