Change 863051 had a related patch set uploaded (by MusikAnimal; author: MusikAnimal):

[mediawiki/extensions/Phonos@master] Add a minimum file size and show an error if a file less than it

https://gerrit.wikimedia.org/r/863051

See what others think of my solution. I got to coding an realized that we can probably safely deduce audio is empty (i.e. created but < 0:01 in length) solely by going by the byte length of the raw MP3 data, and not have to first create the file. This avoids unnecessary operations in Swift. An alternative is do like Extension:Score and use a script to get the length of the audio and go by that, but this would require a trip through Shellbox which we want to avoid if possible.

I tested many short words like "a", "hi", etc., and all are well over the minimum size I went with of 1200 bytes. For now, I've only done this for the Google engine since I know others (eSpeak in particular) generates shorter audio files.

For comparison, the first one here is empty and 1,056 bytes long.

Change 863051 merged by jenkins-bot:

[mediawiki/extensions/Phonos@master] Add a minimum file size and show an error if a file less than it

https://gerrit.wikimedia.org/r/863051

QA notes: Examples of IPA that generate empty audio:

• <phonos lang="en" ipa="ˈkɑːtɑːr, kəˈtɑːr" />
• <phonos lang="ar" ipa="Ķuḍā'ī" />
• <phonos lang="ar" ipa="foobar" />

Hopefully it's easy to create more examples. Basically just refrain from using the text parameter and give it bogus IPA (noting the first two examples above are not actually bogus, though!), it will often fail.

Another thing to be aware of is it's possible (but as-of-yet unproven) that actual playable files are 1200 bytes or smaller. In my testing, I looked at a lot of single-syllable words such as "a", "hi", etc., and none ever seemed to be close to 1200 bytes. So QA'ing might also involve trying to find would-be legitimate Phonos files that never get stored because it's so small.

In T324239#8443393, @MusikAnimal wrote:

Another thing to be aware of is it's possible (but as-of-yet unproven) that actual playable files are 1200 bytes or smaller. In my testing, I looked at a lot of single-syllable words such as "a", "hi", etc., and none ever seemed to be close to 1200 bytes. So QA'ing might also involve trying to find would-be legitimate Phonos files that never get stored because it's so small.

The smallest I have found so far is 1440 bytes. Here is an example <phonos ipa="ɑ̃" lang=fr-ca />.

The largest "empty" file I have found so far is 1152 bytes (testing on the commit before this patch).

I extracted all(?) the IPA phonemes from https://cloud.google.com/text-to-speech/docs/phonemes and created a phonos tag for each (P42430).

@MusikAnimal Oh, just to clarify, we always get an mp3 from google? Per T319379. I am assuming that if we did get a WAV from google it would be bigger (even if empty?)

In T324239#8450981, @dom_walden wrote:

@MusikAnimal Oh, just to clarify, we always get an mp3 from google? Per T319379. I am assuming that if we did get a WAV from google it would be bigger (even if empty?)

Correct, we always get MP3 from Google. For now, there's no minimum on the size of the files for the other engines, since I didn't bother to figure out what the appropriate value should be. eSpeak for instance makes very short audio files, so the threshold would need to be lower for it.