simple-ehm

A simple tool for a simple task: remove filler sounds ("ehm") from pre-recorded speeches. AI powered. Istruzioni in italiano in fondo al documento.

Quick start

pip install -r requirements.txt
python simple_ehm-runnable.py /path/to/video.mp4 --srt --fastcut
# GUI (same folder): python simple_ehm_gui.py

Output video + .srt debug track are written next to the original file.

Features (v2.x)

• Modern stack: Python 3.10–3.13, TensorFlow 2.16–2.21 / Keras 3, Windows + Linux.
• NVIDIA GPU auto: --device auto (default) uses the GPU when TensorFlow sees one (memory growth + mixed precision), else CPU. --device cpu|gpu to force. --encoder auto (default) uses h264_nvenc when your ffmpeg supports it, else libx264 (ignored with --fastcut). --list-devices prints what your PC can use (TF GPUs + NVENC) and the recommended combo. Note: native Windows TF is CPU-only; for GPU inference use Linux + CUDA or WSL2 (NVENC encoding works on Windows too).
• Auto language: IT/EN picked from the PC (Windows API / POSIX locale), override with --lang it|en or SIMPLE_EHM_LANG. GUI language menu included. Strings live in langs/*.json – add a file to add a language.
• Models from GitHub: first run downloads model.keras only if the bundled one is missing or older than the release manifest.json (SIMPLE_EHM_REPO, default morrolinux/simple-ehm). Override: --model, --model-url, SIMPLE_EHM_MODEL_URL, --no-download. Program update check included (release tag vs version.py).
• Resume merges: --concat-only re-runs only the final ffmpeg concat from tmp/inputs.txt (no re-analysis).
• GUI: simple_ehm_gui.py (tkinter, stdlib only): dark/light theme, IT/EN switch, progress bar with phase + ETA, live log, Stop (kills ffmpeg too), "Open folder" when done. Advanced options (CRF, FPS, window sizes, model, training data, analysis only, concat only) in a collapsible section. Settings are remembered between runs. python simple_ehm_gui.py video.mp4 pre-fills the file.

Binaries

• Windows: build_windows.ps1 → dist/simple-ehm.exe (≈3.5 GB with TF), dist/simple-ehm-gui.exe (≈16 MB). dist/ is not tracked in git: binaries are published as GitHub Release assets together with SHA256SUMS.txt.
• Before running: verify the hashes as explained in VERIFY.md (SHA256 + Defender/VirusTotal + sandbox).
• Linux: build_linux.sh → dist/simple-ehm, dist/simple-ehm-gui.
• Debian: build_deb.sh → simple-ehm_<ver>_amd64.deb (sudo apt install ./simple-ehm_*_amd64.deb, also shows up in the app menu). PyInstaller specs: simple-ehm-cli.spec, simple-ehm-gui.spec.

Publishing models (maintainers)

Attach to a GitHub Release: model.keras + manifest.json (models/manifest.json is the template: version, url, sha256, size). Clients download only when manifest.version != model.keras.version.

Docker: docker build -t simple-ehm ., then ./convert.sh /path/to/video [args...] (input files must be inside this folder for the bind-mount).

ITA

Un semplice strumento per un semplice compito: rimuovere gli "ehm" dai discorsi pre-registrati.

pip install -r requirements.txt
python simple_ehm-runnable.py /percorso/al/video.mp4 --srt --fastcut
# GUI: python simple_ehm_gui.py

• GPU NVIDIA automatica: --device auto usa la GPU se TensorFlow la vede, altrimenti CPU. --encoder auto usa h264_nvenc se il tuo ffmpeg lo supporta, altrimenti libx264. --list-devices dice cosa può usare il PC. Su Windows nativo inferenza solo CPU; per la GPU serve Linux + CUDA o WSL2 (l'encoding NVENC funziona anche su Windows).
• Lingua automatica: IT/EN dal PC (API Windows / locale Linux), --lang it|en o SIMPLE_EHM_LANG per forzare.
• Modelli da GitHub: al primo avvio scarica model.keras solo se manca o è più vecchio del manifest.json della release.
• GUI: python simple_ehm_gui.py: tema scuro/chiaro, lingua IT/EN, barra di avanzamento, registro, Ferma, "Apri cartella" e opzioni avanzate.
• Binary: build_windows.ps1 (exe), build_linux.sh + build_deb.sh (.deb).

Contribuire al modello / dataset

Vedi le istruzioni originali nel notebook simple_ehm.ipynb e trainingset/: clip WAV pcm_s16le mono 16 kHz per classe (silence, speech, ehm), estraibili con --generate-training-data.