Skip to content

Measurement method

Equipment and layout

  • Microphone: Dayton iMM-6S with its calibration file (its calibration file is in data/mic/ for that one microphone only; load your own in REW and in analysis). Placed at the listening seat, ear height, opposite the centre speaker.
  • Capture device: a SPACETOUCH USB audio adapter (ALSA card 1), separate from the playback interface. Two devices means two clocks: all timing methods here remove the drift.
  • Level: ./scripts/rec.sh gain -18dB set the capture level once. Do not change it during a session.

Recording with rec.sh

./scripts/rec.sh list                                  # find the capture device (not a .monitor)
export MIC=alsa_input.usb-SPACETOUCH_USB_Audio_000000000-00.analog-stereo
./scripts/rec.sh gain -18dB                            # once per session
./scripts/rec.sh NAME 45                               # records ./recordings/NAME.wav (48 kHz stereo, 24-bit)
./scripts/rec.sh check recordings/NAME.wav             # peak and clipping check

Start the clip or test signal about 3 s after the recording starts so the first seconds are a noise floor.

Order for an A/B of presets

Record "direct" (no DSP) first, then each preset, then "direct" again at the end: the two direct recordings show whether anything drifted (mic position, room noise, level). Keep playback level, mic position and capture gain identical.

Clipping

If check says the peak is above about -6 dBFS, treat the recording as rough. Loud scenes clipped at the mic; lowering the playback level by roughly 10 dB gives a clean recording and the change in dB between presets is still valid.

What each test is for

Test Tool Answers
Frequency response at the seat REW sweeps (data/rew/) Where the room modes and nulls are; fit cut filters
Preset comparison rec.sh recordings of real film clips, band-energy comparison against "direct" Did the preset change what was meant to change, by how much
Arrival time and level per speaker make-delaytest.py + analyze-delaytest.py Delays and trims
Imaging by ear make-bedtest.py Channel order, phantom centre, pans, click alignment, polarity pairs, dialogue

The bed test

python3 scripts/make-bedtest.py bedtest-5.1.flac writes a 163 s 5.1(side) FLAC and prints the start time of each part: T1 channel order, T2 phantom versus real centre, T3 front pan, T4 around you, T5 click alignment, T6 polarity pairs, T7 dialogue-like noise. Play it with mpv-atmos so it goes through the same chain as films.

Proof of the alignment (still open)

Record the first 14 seconds twice, once with the alignment off and once on:

./scripts/set-eq.py --set=drone2                 # alignment off
mpv-atmos --pause --end=14 bedtest-5.1.flac      # second terminal; then ./scripts/rec.sh bed-off 20, press space about 3 s in
./scripts/set-eq.py --set=drone2 --align=bed     # alignment on
mpv-atmos --pause --end=14 bedtest-5.1.flac      # then ./scripts/rec.sh bed-on 20

The per-speaker level difference between the two recordings should match the trims: SL -4.4, FR -1.9, FC -0.8, SR -0.7 dB. (With the router installed no manual routing step is needed.)

Sweep tests with CamillaDSP in the path (2026-10-23)

REW sends its sweep straight to the sound card, so it cannot measure through CamillaDSP. tools/sweeptest/ and tools/subtest/ play log sweeps on the eight channels (and pairs) through play-through-camilla.sh, record on the mic, and deconvolve with sync chirps as the timing reference. analyze-sweeptest.py reports per-speaker arrival and tone, and the heights-versus-bed fit; --heights-on checks a recording made with the heights processing active. analyze-subtest.py reports the sub's range, compression and distortion, and a rattle hunt with stepped tones. Both tolerate a late-started or early-stopped recording (they place the file by the sync chirps). Each folder has a README with the commands.

Dialogue and reflection tests (2026-10-23)

tools/dialoguetest/ (centre-speaker response, real versus phantom centre, C50, RT60, STI, background noise) and tools/reflecttest/ (early-reflection energy-time curves from the existing sweep test, with a panels-on versus panels-off comparison and an image-source label per surface) are built and checked against simulated or existing recordings; see their READMEs. The reflection comparison's repeat-run noise is about +-3 dB per time slice.

Real-scene test (2026-11-15)

tools/realscenetest/ compares the speech band's balance against the rest of the sound for a real film scene, as authored and as recorded at the seat (microphone calibration applied). It tells bass the room adds apart from other causes of a dip in loud passages. It cannot detect a loss that affects only the dialogue: its self-check turns the source's dialogue down 12 dB in the loud passages and the readings move by 1 dB or less on the author's scene. On that scene about 5 dB of the dip in loud passages was bass; above 200 Hz the balance held. See its README for the numbers and limits.

Things that bit us

  • REW sweeps were recorded with no timing reference on a separate capture device: do not read absolute arrival times from them.
  • Mic and output share no clock: always use a drift-removing fit.
  • A mic recording of a very loud scene is only a rough comparison (ceiling clipping).
  • The source of a hot trailer can itself exceed full scale (a few samples on FR/SR in one clip): that is the file, not the chain.
  • Repositioning the mic between runs changes single 1/6-octave bins by 3-7 dB (and the bed mean by 1-2 dB): compare band averages and reference channels, not single bins.
  • The .mdat files of REW 5 are Java serialisations; javaobj-py3 reads them. A 45-58 dB "response" under 90 Hz in the height sweeps was noise, not output.