← All tutorials

Lab 2 — Analyse output with Wave Compare

Take two renders of the same signal — a float reference and a fixed-point port — into Mantissa's Wave Compare mode. Run wav-diff, read the MLD / MAD / SSNR report per channel, and jump from the worst-mismatch marker straight to the sample that broke. 20–30 minutes.

What you'll need

• Mantissa (you have it — that's how you got here).
• Two .wav files of the same signal — a float reference and a fixed-point port. Either the out_float.wav / out_fixed.wav you produced in Lab 1, or the sample pair below.
• No compiler and no IDE this time — this lab lives entirely inside the app.

1. Get the two files

If you finished Lab 1, you already have its two outputs — use out_float.wav as the reference and out_fixed.wav as the candidate.

Otherwise, grab the sample pair — a float reference and a fixed-point port carrying a deliberate, audible regression:

Download the sample WAV pair

2. Load the pair into Wave Compare

Switch to Wave Compare mode

1. Click the button above to open the Wave Compare mode. (On the public web this navigates to the Wave Compare info page — the mode itself lives in the desktop app.)
2. Load the float .wav as the reference, and the fixed .wav as the candidate. Wave Compare overlays them on one shared time axis.
3. The mode accepts a second fixed candidate too (the optional Fixed2 row) — one is enough for this lab, but that's how you'd A/B two ports against the same reference.

3. Run wav-diff

Press Compare. Wave Compare runs wav-diff per channel and fills in the report. For each channel you get four numbers:

• MLD — maximum loudness difference: the peak perceptual-loudness mismatch under the BS.1387 / 3GPP TR 26.843 §5.3.3 PEAQ ear-model. Picks up the kind of error a human listener would notice, not just bit-for-bit deltas.
• MAD — maximum absolute difference: the single worst sample-by-sample error, in raw amplitude.
• min SSNR — minimum segmental SNR: the signal is split into 20 ms windows and a per-segment SNR (in dB) is computed; the report shows the worst (lowest) one. Catches localized blow-ups that a global RMS error would average away.
• The sample index of each worst-mismatch point — one per metric, so you can seek directly to it.

A stereo file reports each channel separately — a regression that hits only one channel shows up immediately.

4. Jump to the worst mismatch

1. Click the MLD marker on the plot. Every overlay seeks to that sample at once.
2. Zoom in. You can see exactly where the fixed port left the reference — a clipped peak, a stuck sample, a slow drift.
3. Note the shape of the divergence: a lone spike reads differently from a sustained gap (see below).

5. Read the verdict

Each metric flags a different kind of failure — read them together, not just the worst number:

• MLD flags errors a human listener would notice. It's a perceptual-loudness diff under BS.1387's ear-model, so a high MLD means audible divergence in critical bands — even if the raw amplitude error is tiny.
• MAD flags large raw-amplitude errors. It says nothing about audibility: a big MAD can land in an inaudible frequency range or during a silent passage and never be heard.
• min SSNR flags time-localized blowups. The signal is split into 20 ms segments, each segment's SNR is computed, and the worst one is reported. A much-lower min SSNR than the average SNR means one specific window broke — usually a transient where a fixed-point op overflowed, saturated, or quantized into noise.

Reading the three together:

• All three healthy — small MLD, small MAD, high min SSNR. The port is doing its job; whatever error is left is uniform broadband quantization noise at half-LSB level.
• MLD high, MAD modest, min SSNR fine — the error isn't huge in raw amplitude but the ear-model says it lands in an audible band. Suspect quantization in a critical band, or a small bias the ear is sensitive to.
• MAD large, MLD modest — big raw errors the ear doesn't object to. Suspect errors in inaudible high frequencies, or large errors during silent passages.
• min SSNR collapses, MLD/MAD only moderate — a localized glitch. One or a few 20 ms windows are far worse than the rest: a saturation or overflow on a loud peak. That's the bug Lab 1 hunts. Click the SSNR marker on the plot to jump to the offending segment.
• MAD lifted across the board, MLD moderate, SSNR steady — a systemic fault: a wrong Q-factor, a scaling error, or a DC offset from a truncated stage (step 5). The mismatch is uniform, so no single segment stands out — the average ratio is what suffers.

That diagnosis — perceptual vs. amplitude, localized vs. systemic — is what points you back at the right fix: a single bad operation, or a format decision taken too early.

And that's the loop end to end — pick a format, port the arithmetic, debug it against the float reference, and verify the output. The same five tools and two modes carry any fixed-point port, however large.

Download the sample WAV pair · ← All tutorials