Claude is deaf, let it hear your video with earshot

Give Claude Code the repo link and ask it to install it:

Install this repo: https://github.com/Leonxlnx/earshot

1. Follow the steps in the README's "Install" section; use its own virtual environment (.venv).
2. Install torch with the CPU build, as the README says.
3. Before installing anything, tell me what you're about to do.
4. When done, run "earshot doctor --load" and tell me the result.

The install clones the repo, creates a virtual environment, and installs the CPU build of torch and earshot. On Windows the virtual environment's path is .venv/Scripts/python; on Mac and Linux it's .venv/bin/python.

earshot doctor --load

This checks the install, the models and the CPU settings. On the first run it downloads the two models (~1.1 GB). After that it reads them from disk every time and never goes online.

If the earshot command isn't found, the virtual environment isn't active. Telling Claude "run earshot from its virtual environment" is enough.

In the folder with your video:

earshot critique video.mp4 --fast

critique compares your video with a profile built from 7 professional launch films. It says whether each metric is in range and puts a "fix this first" list at the top.

--fast skips two heavy checks, whether the sound is a real recording or synthesized and whether there's music, so the result comes in seconds instead of minutes. Run it once without --fast before you publish.

For my own video the summary was: 16 of 21 metrics in range, 1 critical and 1 major issue. The critical one was loudness:

• My video: −23.9 LUFS
• Web and YouTube target: around −14 LUFS
• Difference: ~10 dB. Viewers turn their phone up to hear the video, then down again for the next one.

Peaks reached −4.4 dB, so the audio was a whisper in one place and a bang in another.

What is LUFS? The average loudness as the ear perceives it. The closer the number is to zero, the louder: −14 is louder than −24.

The report is full of technical terms. Instead of working through all of them, I copied the output and pasted it into Claude Code:

Can you explain this technical analysis simply enough for me to understand? Start with the 3 most important problems and, for each, tell me in one sentence what I should do.
<paste the earshot critique output here>

What Claude pulled out for me: the audio is very quiet, there are too many emphasis sounds, the audio is fully mono, there's almost no bass, and the mids are a bit harsh. It also reminded me that two things aren't measured in --fast mode.

Mono audio isn't a problem for me. I make content for phones, and a phone speaker plays more or less as one channel anyway. You don't have to fix every finding; you decide which ones matter to you.

Loudness was the biggest problem, and it's also the easiest to fix. Ask Claude to bring the audio to −14 LUFS with ffmpeg:

Bring the audio of video.mp4 to −14 LUFS with ffmpeg, with true peak no higher than −1.5 dB. Don't touch or re-encode the picture. Name the result video_14lufs.mp4. When done, measure the new loudness with earshot measure and tell me.

earshot measure only measures and loads no models, so it takes seconds. That's how you see the fix worked.

The audio of this guide's own video was fixed the same way: from −21.2 LUFS to −14.1.

Every earshot command starts a new process and loads the models from scratch: 20-40 seconds each time. If you'll measure a lot on one video, connect earshot as an MCP server. The models load once and stay loaded:

claude mcp add earshot --scope user -e EARSHOT_THREADS=3 -- /path/to/earshot/.venv/bin/python -m earshot mcp

Replace the path with your own earshot folder. On Windows it ends in .venv/Scripts/python. On a slow machine, start Claude Code with MCP_TIMEOUT=60000.