earshot is an MIT-licensed toolkit that gives coding agents ears. It runs on a CPU as a command-line tool, a Python API and an MCP server, and it describes, compares, searches and critiques sound.
It combines two open audio models (CLAP and AST) with objective measurements. Its accuracy numbers come from the author’s own checks, not from an independent test.
The short answer
- Describe and match: it tags a sound and scores how well it fits a text description: strong, plausible, weak or no match.
- Search and rank: it ranks files against a description and finds the best moments in long recordings.
- Measurement: LUFS, true peak and spectrum, without models.
- Sound generation: “Forge” makes sounds from code, with a bank of 201 code-defined effects.
- Requirements: CPU, Python 3.10+, ffmpeg, and about 1.1 GB of models downloaded at first use.
How accurate is earshot, according to its author?
The README reports these results on the author’s own library:
- Matching: on 125 held-out real recordings, a “strong” verdict was the right kind of sound 98% of the time and the exact variant 61% of the time. The page does not call these recordings blinded.
- Real versus synthesized: ROC-AUC 0.92. Firm verdicts misclassify 6.7% of real and 5.5% of synthesized clips, and about a fifth get “uncertain”.
- Forge blind test: a listener picked the made sound 8 of 15 times, which the author calls a coin flip.
What this means: the tool is good at telling you what kind of sound you have (98%), weaker at naming the exact variant (61%), and its synthesized sounds were hard to tell from real ones in one small test. All of it is self-reported.
What can earshot not do?
The README warns that Forge optimizes earshot’s own models. Those can be satisfied without the sound actually improving, so it recommends model-free measurements and human listening.
It also says earshot cannot judge these things:
- taste and emotion
- provenance and licensing
- speech content
- content above 14 kHz
Tags inside busy mixes are only hints. It is not a speech-to-text tool and makes no claim about Arabic.
What should you check before using it?
- Model licenses: the code is MIT, but CLAP is Apache-2.0 and AST is BSD-3-Clause under their own authors’ terms.
- Speed: each command-line call reloads the models in roughly 20 to 40 seconds. The MCP server keeps them loaded.
Who is this for, and what should you do next?
- Good fit: developers who let an agent build or tune sound effects and want the agent to check loudness and rough sound type.
- Poor fit: anyone who needs transcripts, taste judgments or licensing checks.
- Use the MCP server if you need many calls, since it keeps the models loaded.
- Read the license terms of CLAP and AST before you ship anything that uses them.
- Pair its measurements with human listening, as the README advises.
Questions and answers about earshot
Does earshot transcribe speech?
No. The author lists speech content among the things it cannot judge.
Does it need a GPU?
No. The README says it runs on CPU only.
Can I trust the 98% figure?
Treat it as the author’s own check. It means the right kind of sound on a “strong” verdict, not independent accuracy.
Can I use it commercially?
The code is MIT. The models carry their own licenses (Apache-2.0 for CLAP, BSD-3-Clause for AST), so check those too.
Terms in plain language
- MCP server: a small service that lets an AI agent call a tool.
- LUFS: a unit for how loud audio sounds.
- True peak: the highest level the audio signal reaches.
- ROC-AUC: a score from 0 to 1 for how well a test separates two groups, such as real and synthesized.
Sources
Facts read on 21 October 2026.
