Ears for coding agents: let AI hear, review and make sound. CPU-only, open source.
What earshot can and cannot judge
earshot gives an agent measured evidence about sound, not taste. Read this before trusting a verdict.
Trustworthy
- Measurements: loudness (LUFS, true peak, LRA), timing (onsets, offsets to cuts and cue frames, to the millisecond), levels, spectrum, envelope, clipping, silence, tempo of steady music. These are meters, not opinions.
- "Is this the kind of sound I asked for?" on single, exposed sounds (a key press, a whoosh, rain, a glass ting): a "strong" match is very reliable on held-out recordings; "plausible" and "weak" are hints worth a second check.
- Picking among candidates (
rank) and finding the clean moment (find) in real recordings. - Obvious code-synthesized sources (noise risers, simple FM bells, steady test tones): flagged with ~92% precision when the verdict is firm (6.7% of real clips get called synthesized, 5.5% of synthesized clips real). Bare beeps are less reliable than that list suggests: with the beep family held out, beeps were among the most-missed types (accuracy 0.71).
- Whether sounds land on the picture: offsets of hard hits to cuts / cue frames (median 0 ms on Clawd, 96% within a frame).
Hints only
- Tags of events inside a busy mix (what each hit "is"): right about half the time for exposed events, a quarter for events under the music. Times and levels are right; the label is a guess.
- Music genre and mood: often wrong (an upbeat track's top mood was a dark one 18% of the time); describe keeps them apart as unvalidated hints. Instruments, rain, birds and voice are usually right.
- "Uncertain" from realness on designed library sounds ("magic" chimes/celesta, "motion" whooshes) and on real electronic music: they sit between real and synthesized. Read it as "sounds designed", not "fake".
- The launch scorecard ranges come from 7 films the author picked. They describe that style (modern SaaS / AI
launch films, kinetic type, UI demos), not every good film. A cinematic short or a quiet piano piece will break
rules that do not apply to it. Build your own profile from films you admire (
earshot profile). - The "cuts without a hit" list (timeline, sheet, critique tips): soft build-ups (whooshes, risers, swells) can still be missed - on Clawd both cues it named were sounded whooshes. Look at the sheet before adding a sound.
- critique's "onset on visual events" share counts only new attacks within +-100 ms (the definition the reference ranges were measured with), so a whoosh swelling through a cut does not count; its fixes and tips name only the cuts with no sound at all (timeline's matcher). critique's stop-downs (short drop-outs before a hit, per minute) and timeline's gaps (silences of any kind) are different counts.
- realness verdicts depend on length: a sound up to 6 s is judged with the clip thresholds (lo 0.17), a longer file
by the mean of its 4 s windows with the weakly calibrated file thresholds (lo 0.40), so a lower p_real can get the
milder verdict. Judge the trim you will actually use (
--start / --end), not a multi-take file. - AudioSet labels on one-shots under ~0.5 s: often 'unsure' or a near class (Zipper for a swish, Finger snapping for a click). Treat those as neutral; only an outright different class (Speech, Music, Engine) is a red flag.
Cannot judge
- Taste and emotion: whether a sound is beautiful, whether the music fits the brand, whether a moment feels satisfying. earshot can say "a strong match for a deep thocky key press, 3 ms attack, lands 4 ms after the cut"; it cannot say the film feels premium.
- A code-made layer hidden under a louder real mix: usually missed. Judge source files and stems, not only the final mix.
- Provenance: realness hears synthesis, not licences. A professionally made, licensed synthetic UI sound can read "likely code-synthesized"; code-made noise beds that imitate nature (rain, wind) or telephone tones can read real.
- Anything above 14 kHz from the models (CLAP's front end stops there; the AudioSet tagger at 8 kHz). Use
measure(bandwidth, bands) for air, hiss and codec cut-offs. - Speech content: no transcription; voice is detected, words are not.
- Licences: earshot never decides whether a file may be used. The real-sound skill's
audit.pychecks manifests; the licence itself must be read on the item's page.
Known engineering limits
- Cold start: every CLI call is a fresh process that loads the models (~20-40 s on a busy laptop), even for a cached
file; the MCP server or one Python script (
from earshot import api) loads them once.doctorwithout--loadloads nothing heavy. - bf16 is used only on CPUs with native bf16 (AVX512-BF16 / AMX); elsewhere earshot runs fp32 (bf16 emulation was
measured 16x slower). The bundled calibration was measured with bf16 embeddings (fp32 agrees to cosine >= 0.9993 on
12 real clips,
eval/semantics/precision.json). - Caches are keyed on path, size, modification time and 16 sampled 4 KB blocks of the file: a re-render that differs
only between those blocks and keeps the old size and timestamp would still read stale. Use
timeline --freshor touch the file when in doubt. - The models were evaluated on one sound library (Freesound / Mixkit picks for one film), 7 launch films and three synthesis code bases. Other domains (orchestral recording sessions, field recordings in other climates, other synth engines) are untested.
- The optional Gemini critique has not been validated against real answers (it needs the user's key and consent to upload).
- Cut detection marks big visual changes (a blinking cursor counts); slow crossfades are missed.