erm
Local CLI that strips disfluencies (um, uh, er, erm, ah, hmm, mhm,
mm, uh-huh, plus any-length elongations like ummmm / uhhhhh) from
recordings of English speech.
It uses faster-whisper (running
the large-v3 Whisper model by default — override with --model) for
word-level timestamps, three audio-domain detectors that catch fillers Whisper
hides, and ffmpeg for the cuts. Each splice is snapped to a local energy
minimum and zero-crossing, optionally crossfaded with a length that scales
with the cut size, and laid over a constant looped sample of the recording's
own room tone so the noise floor stays uniform across edits.
Full docs at dougcalobrisi.github.io/erm
(source in docs/): usage guides for getting good
results — tuning & workflow, recipes,
troubleshooting — plus maintainer-facing design docs
on the detection passes, render pipeline, denoise/room-tone, and transcription.
Install
Requires Python 3.11+ and ffmpeg / ffprobe on PATH.
python3.13 -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
Transcription device (GPU vs CPU)
Transcription runs on CPU by default and needs no extra setup. If you have an
NVIDIA GPU, faster-whisper can use it — but only when the CUDA runtime libraries
(libcublas, libcudnn) are installed. A machine with an NVIDIA GPU and driver
but no CUDA runtime is the common case that produces:
RuntimeError: Library libcublas.so.12 is not found or cannot be loaded
erm handles this automatically: with the default --device auto, if the GPU
can't be loaded it prints a warning and falls back to CPU, so transcription
still completes. You have two ways to make it explicit:
-
Force CPU (no warning, skips the GPU probe): erm input.wav --device cpu
-
Enable the GPU by installing the CUDA wheels into the same environment:
pip install nvidia-cublas-cu12 nvidia-cudnn-cu12
faster-whisper's CUDA backend needs CUDA 12 / cuDNN 9. See the
faster-whisper GPU notes
for details.
Usage
# Remove fillers; output and cut-list paths are auto-generated next to the input.
erm input.wav
# Specify output explicitly.
erm input.wav -o cleaned.wav
# Inspect what would be cut without rendering.
erm input.wav --dry-run
# Validate a rendered output against its source.
erm validate input.wav cleaned.wav --cuts cuts.json
# Pull just the cleaned AUDIO out of a video (default — no picture rendered).
erm talk.mp4 # -> talk-cleaned-*.wav
# Keep the PICTURE: render a synced video output (container inferred from input).
erm talk.mp4 --video # -> talk-cleaned-*.mp4, audio + video in sync
When -o / --json are omitted, output paths are written next to the input as
{stem}-cleaned-{YYYYMMDD-HHMMSS}.wav and {stem}-cuts-{YYYYMMDD-HHMMSS}.json.
Video
A video input works like any other recording, and by default erm emits the
cleaned audio only (.wav) — the common "pull the audio out of this video"
case, and identical to how erm has always behaved. Pass --video to render
the picture too:
- The output container is inferred from the input (mp4→mp4, mov→mov…); an
-o extension always overrides.
- Audio and video stay in sync by construction (within ~1 frame): both
streams are rendered from the same edit timeline with the same frame-snapped
crossfades, and the picture is conformed to the audio's exact length.
--mode silence stream-copies the picture untouched (lossless, frame-exact).
--mode remove re-encodes the spliced picture; --video-splice crossfade
(default) dissolves each splice, --video-splice cut makes hard jump cuts.--min-gap-ms with --video plays the removed footage through the
injected pause (muted) rather than freezing the frame.- Audio is stored losslessly where the container allows (PCM in mov/mkv/avi);
mp4 gets AAC 256k, webm gets Opus 160k.
See Modes and the
video render & A/V sync design doc for the A/V-sync
derivation.
Use inside AI coding agents
erm ships agent guidance so an AI assistant can install, run, and tune it for you.
Claude Code / Cowork (plugin):
/plugin marketplace add dougcalobrisi/erm
/plugin install erm@erm
This adds two skills: erm (install + clean a file, asks which mode/recipe
fits) and erm-tune (diagnoses a bad result and maps the symptom to the
right knob). Ask in plain language — "install erm", "clean up this podcast", "the
splices sound smeared" — and the right skill triggers.
Other agents (Codex, Copilot, OpenCode, Cursor, Gemini CLI, pi.dev):
