A practical A/B checklist for pitch- and tempo-shifted audio

[PitchTempoAndriy]

A louder render can sound more impressive even when the actual change is less clean. Here is a small comparison workflow for anyone checking a pitch shift or stretch before committing it to a mix.

1. Keep the untouched source and import a separate rendered copy into Audacity. Choose the same musical phrase in each version. If the tempo changed, match the phrase, not an identical number of seconds: otherwise one excerpt may contain a drum fill or chorus that the other does not.

2. Make comparison copies of those excerpts. In Effect > Volume and Compression > Loudness Normalization, use the same perceived-loudness target for each. Audacity's -23 LUFS default is a useful quiet starting point for this test, not a required release level. Leave “Normalize stereo channels independently” off so that the comparison preserves each file's left/right balance. Check for clipping; if needed, use a lower target for both.

3. Alternate the excerpts, with only one playing at a time and your playback volume fixed. Peak normalization alone does not guarantee similar perceived loudness. Even LUFS matching is a starting point: tonal changes can still affect how loud something feels.

4. Check one thing per pass: drum attacks for smearing or extra clicks; sustained notes for wavering; consonants for intelligibility; and reverb tails for abrupt cutoffs. Listen beyond the opening, including a busy section and the ending. A short preview cannot establish that the whole render is clean.

Keep the listening copies separate from the mix files. A level adjustment made for comparison is not a decision to change your mix balance. Also, a genuinely different key or tempo will change the musical impression; that alone is not evidence of a processing defect.

Audacity reference: https://manual.audacityteam.org/man/loudness_normalization.html

 

[rob aylestone]

Please forgive me, but I'm not quite sure what your process is actually for? Chopping up before and after processing so you can hear differences? I don't think I've ever done that - or to be honest needed to. I'm very interested in what your system actually does for you in use? What sort of things does it highlight as wrong?

 

[PitchTempoAndriy]

Thanks, Rob. I made a simple listening check sound rather elaborate.

It's meant to help decide whether a pitch/time-stretch render is usable, or whether to try a smaller change or another setting. The excerpts are just disposable listening copies made after processing; choosing the same phrase makes comparisons quicker.

Examples of things to listen for are a drum attack becoming smeared, a held vocal note acquiring a watery wobble, or a phrase ending losing part of its reverb tail. Bringing levels reasonably close helps avoid favouring the louder version.

If a small change already sounds fine in context, a normal listen may be enough. This is a checklist for a render you're unsure about, rather than a necessary new editing method.

 

[TalismanRich]

I don't know what loudness and compression has to do with pitch shifting.

I've done pitch shifts in Reaper and could tell no difference, except that I changed a song from Bb to A. For relatively minor changes, it can be pretty much a wash. Reaper lets you change pitch without changing tempo, and vice versa. It has no effect on volume.

When you start talking about changing 3 or 4 notes, it can show up with obvious artifacts. I tried doing a harmony part that was out of my range. I lowered the track enough that I could reach all the notes, then took that track and brought it up to the original pitch. It sounded bad, a bit of that Chipmunks sound. Even blended in with normal vocals, it didn't work. Turns out that a higher voice isn't just a different pitch.

 

[PitchTempoAndriy]

Thanks, Rich. I wasn't suggesting compression as part of pitch shifting: 'Volume and Compression' is just the Audacity menu containing Loudness Normalization. The level check was meant to avoid a louder comparison winning by default. If source and render already have comparable perceived levels, there's no reason to normalize them for this check.

Your vocal example is useful here: a changed vocal character can be the issue even when volume is unchanged. I should have made that separation clearer in the opening post.