Let a coding agent design sound effects by listening
Leonxlnx built earshot so coding agents can describe and create sounds; they say every sound in a 22-second launch film came from it.
Done with Claude Code
What they did
Leonxlnx built earshot, a free tool that turns audio into written descriptions an agent can read. It scores how close a sound is to a real recording and lines sounds up with video cuts. A feature called forge lets the agent write a recipe for a sound in code. The tool tries many versions, scores them against real recordings, and the agent rewrites the recipe based on feedback.
How it went
The author says every sound in the 22-second launch film was made by forge. In a blind test of five sound types, they picked the made sound 8 of 15 times, about a coin flip. Only the made impact sound was spotted, because it fades too fast.
Worth knowing
The search can learn to please the tool's own scoring without the sound really improving, and the realness check is easy to fool. The author advises trusting plain sound measurements, visual sound charts and a human listening test.
Try it yourself with Claude Code
Using the earshot MCP server, listen to [video or audio file] and give me a timeline of sound events and how well they sync to the picture cuts. Then run a critique against the launch-film profile. List the top fixes with timestamps. For [sound type, e.g. keypress], use earshot forge with references in [folder of real recordings] to make candidate sounds in code. Read the critic's feedback, revise the recipe and iterate. Show me the best few WAVs with their scores and spectrograms before replacing anything in my project.
Swap the highlighted [parts] for your own details.
Source: GitHub · Leonxlnx · Oct 15, 2026
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
