A browser app for tuning any DSP or parametric EQ — car audio or home (speakers, AV receiver, PC EQ) — without a measurement microphone. You play a song through your system, note a clean moment (t1) and the moment that booms or bites (t2). The app compares the two windows of the source audio, ranks the frequencies that surged at t2, and lets you solo each one through your speakers to confirm by ear. The confirmed ones become PEQ entries (frequency, Q, gain) for your DSP.
▶ Use it online: https://emdaitaj.github.io/dsp-no-mic-eq/ · Pull requests are welcome (see Contributing).
Everything runs locally in the browser. Nothing is uploaded and it works offline.
- Online: open https://emdaitaj.github.io/dsp-no-mic-eq/. Your audio files are analysed on your device and never leave it.
- Single file: open
dist/tdsa.html(double-click). Copy it to any laptop or phone. - From source: open
index.html, or serve the folder (python -m http.server 8765) and browse tohttp://localhost:8765. - Press Load demo for a synthetic song with a known boom (B1, 61.7 Hz) and a 2.8 kHz harshness.
Recent Chrome, Edge, Firefox or Safari. WAV, FLAC and MP3 decode everywhere; OGG/M4A where the browser supports them. Files are decoded at their native sample rate, so a 44.1 kHz file keeps 44.1 kHz FFT bins. Hi-res files (88.2–384 kHz) are decoded at 44.1 or 48 kHz: nothing above 20 kHz matters for EQ, and the FFT bins stay fine.
- Connect the computer or phone to your system (USB / optical / AUX / Bluetooth). Keep Output at −6 dB or lower.
Optionally add headphones as a second device under Outputs (desktop browsers), line the two up with the delay, and press
Sto switch between the car or room and the headphones at the same moment of the song. - Play the song in the app. When it booms or bites, press 2; at a clean passage press 1. Your reaction time is subtracted (250 ms, adjustable). Fine-tune either moment with − / + (or ↑ / ↓ in its time field).
- Read the DSP EQ Action Table (top 5 suspects).
- Solo each suspect: the band that reproduces the boom on its own is the culprit. Sweep the dock's frequency while it plays to pin the loudest point.
- EQ A/B loops t2 through the suggested cut; toggle A/B to hear whether it fixes the problem.
- Find draws that band's energy over the whole song and lists its loudest moments. They should boom too.
- + EQ collects confirmed cuts; Copy EQ list gives a per-speaker-way list to type into your DSP software.
- Repeat with 2–3 other songs and Log the suspects. Frequencies that recur in the Resonance Log belong to your system or room, not the music.
| Step | Method |
|---|---|
| Windows | t1 and t2 centred on the timestamps, default 400 ms (50 ms – 10 s). Mid (L+R)/2, Left, Right or Side. |
| PSD | Welch, FFT 16384, Hann, 75 % overlap, mean-detrended segments centred in the window. FFT size 4096–65536 and segment length are adjustable. With a 0.4 s window a 16384-point segment fits once, so the estimate is a single Hann periodogram. Windows ≥ 0.65 s get real Welch averaging. |
| Level scale | dBFS, where a full-scale sine reads 0 dBFS (AES17). Band levels integrate the PSD between band edges (sub-bin bands are interpolated). |
| Views | 31-band ISO ⅓-octave centres 20 Hz – 20 kHz (the bands of a 31-band graphic EQ); 1/12-octave semitone bands centred on equal-tempered notes (A4 = 440 Hz), 19.4 Hz – 19.9 kHz. |
| Curves | Spectrum at t1, spectrum at t2, Δ(f) = dB(f, t2) − dB(f, t1). Optional level-matched Δ removes the median broadband change (use it when t2 is simply louder, e.g. chorus vs verse). |
| Prominence | t2 band level minus the mean of the ±4 neighbouring 1/12-octave bands. |
| Suspect Score | Surge⁺ + Prominence⁺ × gate, where gate = clamp(Surge / 6 dB, 0, 1). A frequency that did not get louder at t2 cannot explain a problem that only appears at t2. Without the gate, steady tonal content that is already present at t1 (sustained chords) outranks the real boom. The literal plain sum is available in Settings. Peaks are local maxima of the t2 1/12-octave spectrum, at least ⅓ octave apart. |
| Centre frequency | Parabolic interpolation of the raw t2 spectrum for narrow (tonal) spikes; the middle of the half-prominence span for broad humps. |
| Q | Width at half prominence on a 1/24-octave-smoothed t2 spectrum. That is the bandwidth definition of a peaking EQ (RBJ), so the value is directly usable as a cut's Q. “≥” marks spikes as narrow as the window can resolve. |
| PEQ start | Q clamped to 0.5 – 8 (adjustable: set your DSP's Q limits in Settings). Cut = −(1.5 + 0.15 × score) dB, limited to −2 … −9 dB. It is a starting value: confirm it with EQ A/B. |
| Way | Sub / Midbass / Midrange / Tweeter from your crossover points (Settings), with a note near a crossover. |
Q values follow the standard RBJ peaking-EQ definition (Q = centre ÷ bandwidth at half the gain in dB). DSPs differ in how they define Q or bandwidth, so check yours.
Verified numerically (node tests/run-tests.js): a full-scale sine reads 0 dBFS ±0.25 dB, Parseval holds to 0.02 %,
ISO 226 matches the standard, a known +12 dB / Q 4 resonance is measured as Q 4.3, and a −6 dB cut with the sweep Q suggester's Q
is −3 dB (half depth) at the marked edges of the range. For a lopsided 180–430 Hz range peaking at 360 Hz, two fitted cuts follow
its shape about twice as closely as one cut, reach the full depth at the peak and cut less outside the range. On the demo track the
B1 boom (61.7 Hz) and the 2.8 kHz resonance (true Q 5, measured Q 5–7) are both ranked in the top 3.
- Audition dock: Solo (8th-order band-pass, loudness-normalised), Tone (sine, ±½-semitone warble, band noise) and EQ A/B (cut or boost-and-sweep). All parameters change live while the loop plays. ½ / ⅓ / ⅙-octave Q presets. Generated sound goes through a safety limiter.
- Find freq / Find Q (dock, EQ A/B mode):
- Find freq steps a +9 dB, Q 6 boost across ±½ octave in ¹⁄₁₂-octave steps, one loop pass per step, up then down. You press Space where the problem sounds worst. Up and down marks are averaged, which cancels your reaction delay.
- Find Q steps the cut through Q 8.6 → 1.4 (two passes each). You press Space at the first one that sounds fixed.
- Steps are scheduled on the audio clock, so every repetition of the problem sound plays at exactly one setting.
- Speaker side (header): L+R, or a mono mix on the left only / right only speakers, to hear which side's reflections cause a problem and EQ that side only. The mix beside it plays stereo, mono (L+R)/2, or only the left or right channel; with L+R that one signal plays on both sides, so headphones put it in both ears, the way a single car speaker is heard with both ears.
- Two output devices (header, Outputs): everything also plays on a second device B (headphones, car, a second
sound card) next to the main output A. A delay of ±1000 ms (positive delays B, negative delays A) lines the two up, and
A ⇄ B or
S(changeable) switches between A+B, A only and B only. Desktop Chrome, Edge and Firefox; not on iPhone or Android. The browser may ask for the microphone: that is only how it reveals device names, and nothing is recorded. - Spectrogram of the whole song with the ranked suspects drawn as numbered lines.
- Find in song: band-energy timeline plus clickable hot spots, to confirm a suspect at other moments. Use Find in the Action Table, or
right-click any band in the spectrum or Δ chart (or press
Fon a focused chart). - t1 / t2 fine-tuning: − / + beside each timestamp move the window by an adjustable step (default 100 ms, Shift = ×10);
↑/↓in a time field do the same. A playing loop of that window follows it, and Loop t1 / t2 stops its loop when pressed again. - EQ Builder: peaking and shelf bands per speaker way, combined response plot, real-time preview on the track (Play through EQ), EQ list text, CSV and session save/load. Shown as an overlay on the Δ chart and as an optional "t2 after EQ" curve.
- Resonance Log: clusters suspects from many songs and moments (plus sweep marks). Recurring frequencies are flagged as problems of your system or room. Kept in the browser between sessions.
- Sweep Hunter: slow log sine sweep. Press Space (boom), R (rattle) or D (dip); reaction time is compensated and up/down passes are averaged to cancel the remaining bias. Includes a "zoom sweep" around a mark.
The strip shows each mark and harsh range with its frequency; click the strip to jump the sweep to that frequency (or start it there).
- Q suggester: press
1where a harshness starts,2where it is worst and3where it stops (or type the three frequencies). The start and end are taken as the cut's half-gain edges (RBJ Q = centre ÷ bandwidth), so Q = √(lo·hi) ÷ (hi − lo). A peaking filter is symmetric on a log axis, so for a lopsided peak the main suggestion is the least-squares fit (centre = geometric mean of the three marks, full marked width), with a narrow and a wide alternative centred on the peak. Q is kept inside the DSP limits in Settings. Each range has a Depth (default 4 dB, set when typing a range or later in the table) that sets every suggested cut's gain. - Split into 2 or 3 cuts: for an off-centre peak, the range's shape (its Depth at the peak, half that at the start and end, each side as wide as marked, easing to nothing within ⅓ octave past each end) is least-squares fitted with several peaking cuts: one on the peak and others over the long side, kept narrow enough not to reach far outside the range. Every suggestion shows its depth at the peak and what it still cuts ⅓ octave past each end; + EQ all adds the cuts to the EQ Builder. E.g. 10 dB from 180 to 430 Hz, worst at 360 Hz: one cut (303 Hz Q 1.1) reaches −8.6 dB at the peak and cuts −2.5 / −3.7 dB at 143 / 542 Hz; 239 Hz Q 1.9 −6.6 dB + 349 Hz Q 2.9 −8.7 dB reach −10.3 dB with −1.6 / −1.8 dB outside.
- Hearing check: for the sweep range (and each marked range) the ISO 226 contour at your listening level (~60/70/80 phon) shows how much more or less sensitive your ear is in each part. A constant-level 40–500 Hz sweep sounds ~28 dB louder at the top at ~70 phon. It warns when a marked peak sits where your ear is most sensitive.
- Equal loudness: turns the sweep down by exactly how much more sensitive your ear is than at the range's least sensitive frequency, so the whole sweep sounds equally loud (ISO 226 at your volume). It never boosts: the least sensitive part plays at the set level. It can be switched mid-sweep (the change fades over ~0.2 s) to compare. Ranges marked with it skip the hearing caution.
- Q suggester: press
- Ear-Match EQ: third-octave band-noise loudness matching against a reference band, converted with ISO 226:2003 equal-loudness contours into an estimated response and a graphic-EQ correction (flat, tilt or car / room target). Rough (±3 dB), but useful for overall balance.
Space play/pause (marks while Find freq / Find Q or Sweep Hunter runs) · 1/2 set t1/t2 at the playhead (while a sweep runs: 1/2/3 mark a harsh range's start / peak / end) ·
↑/↓ in a t1/t2 field nudge it · A alternate t1/t2 loops · S switch output device A ⇄ B (changeable under Outputs) · Enter analyse · Esc stop all sound · F on a focused chart finds that band in the song ·
+/−/0 zoom · ←/→ nudge the playhead (Shift = 5 s) · O open file · Ctrl+wheel zooms the waveform.
Start quiet. Pure tones and narrow noise bands put far more energy into one frequency than music does. Keep sweeps and tones
low, especially above 2 kHz (tweeters) and below 30 Hz (excursion). Esc or Stop all silences everything.
- The app analyses the music file, not the acoustic response of your room or car. It finds what the song puts into the band where you heard the problem, and your ears confirm it. The measured Q describes the music's spectral spike. A room or cabin mode is usually broader, so the suggested Q is capped at 8 and you should refine it by ear.
- A 0.4 s window cannot resolve 1/12-octave detail below about 70 Hz. That region is shaded in the chart, and a hint offers 1.5 s / FFT 65536.
- Ear-Match and the hearing check depend on ISO 226, which describes average young listeners hearing pure tones in a free field.
index.html app shell
css/app.css styles (light/dark, validated colour palette)
js/dsp.js FFT, Welch, band integration, peaks/Q, biquads, ISO 226, synthesis, header sniffing
js/analysis.js t1/t2 differential pipeline and suspect ranking
js/chart.js log-frequency canvas chart
js/waveview.js waveform, spectrogram, t1/t2 windows
js/audio.js playback, EQ preview chain, audition engine, limiter
js/tools.js Resonance Log, Sweep Hunter (with Q suggester and hearing check), Ear-Match
js/app.js UI controller, persistence, shortcuts
js/demo.js synthetic demo track and WAV encoder
tests/run-tests.js numeric self-tests (node tests/run-tests.js)
tools/build-single-file.js bundles everything into dist/tdsa.html
Settings, the EQ list, the Resonance Log, sweep marks and per-song timestamps are stored in the browser's localStorage.
Use Save session… to keep a copy as a file. After changing any source file, run node tests/run-tests.js and
node tools/build-single-file.js to refresh dist/tdsa.html.
Pull requests are welcome! New features, bug fixes, clearer wording, tips for specific DSPs, car or home setups: all of it helps. For a bigger change, opening an issue first to discuss it is a good idea.
Before sending a pull request, run node tests/run-tests.js (all checks should pass). If you changed the app, also run
node tools/build-single-file.js so dist/tdsa.html stays in step with the source. By sending a pull request you agree to
license your contribution under the same terms as the project (below).
Copyright (C) 2026 emdaitaj
TDSA is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. It is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the license for details.
In short: anyone may use, study, change and share it. If you share a changed version, or let other people use a changed version over a network (for example by hosting it on a website), you must make your source code available to them under the same license. Changes you only use yourself do not have to be published.