Description
Let an AI write the code that renders a product demo and the picture comes out fine, but sound is a blind spot. It cannot tell a premium key press from a cheap one, notice that a hit lands 80 ms after the cut, or hear that the music swells in the wrong place. earshot fills that gap and gives a coding agent a working pair of ears.
It turns sound into evidence an agent can reason about: what a file sounds like in words, how well it fits a description, a second-by-second listening transcript of a whole film, and a scorecard against professional launch films. It also works the other way, letting an agent make sound in code: it writes a synthesis recipe, listens to the result, scores it and revises until it is right. Everything runs locally on CPU, with no GPU and no network.
describe: Give it an audio file and it says in words what it sounds like, with the closest alternative readings plus measured loudness and brightness. A camera shutter from the bank, for example, may read as a snare hit or a box lid closing.
match and rank: Judge how well a sound fits a description, or sort a pile of candidate files by it and pick the best.
find: Locate the best moments inside a long recording and return start and end times ready to trim.
timeline and sheet: Turn a whole video into a second-by-second event list, check whether sound lands on the picture cuts, and render an annotated image with spectrogram for agents that read pictures.
critique: Score a launch film against a profile of 7 professional ones, with a timestamped fix list for weak loudness or misplaced hits.
sfx and kit: A bank of 252 sound effects stored as code that render by id in seconds, aimed at a chosen moment in the picture. A kit renders a whole film’s effects with measured sync points, with no models and no network.
forge and fit: forge lets an agent write a synthesis recipe and improve it by listening; fit translates a recording into recipe code with knobs you can vary.
Three ways to use it: A CLI, a Python API and an MCP server. For Claude Code and other agents the MCP server is easiest because the models stay loaded instead of reloading on every call.
MIT licensed: Written in Python, with the source and all 252 sounds in the repository and playable on the project site.
It turns sound into evidence an agent can reason about: what a file sounds like in words, how well it fits a description, a second-by-second listening transcript of a whole film, and a scorecard against professional launch films. It also works the other way, letting an agent make sound in code: it writes a synthesis recipe, listens to the result, scores it and revises until it is right. Everything runs locally on CPU, with no GPU and no network.
Features
describe: Give it an audio file and it says in words what it sounds like, with the closest alternative readings plus measured loudness and brightness. A camera shutter from the bank, for example, may read as a snare hit or a box lid closing.
match and rank: Judge how well a sound fits a description, or sort a pile of candidate files by it and pick the best.
find: Locate the best moments inside a long recording and return start and end times ready to trim.
timeline and sheet: Turn a whole video into a second-by-second event list, check whether sound lands on the picture cuts, and render an annotated image with spectrogram for agents that read pictures.
critique: Score a launch film against a profile of 7 professional ones, with a timestamped fix list for weak loudness or misplaced hits.
sfx and kit: A bank of 252 sound effects stored as code that render by id in seconds, aimed at a chosen moment in the picture. A kit renders a whole film’s effects with measured sync points, with no models and no network.
forge and fit: forge lets an agent write a synthesis recipe and improve it by listening; fit translates a recording into recipe code with knobs you can vary.
Three ways to use it: A CLI, a Python API and an MCP server. For Claude Code and other agents the MCP server is easiest because the models stay loaded instead of reloading on every call.
MIT licensed: Written in Python, with the source and all 252 sounds in the repository and playable on the project site.
