I'm a programmer without any real audio engineering experience playing around with procedural audio generation. I need to create a notification sound that sounds decent/not-horrible over a potentially low-quality conference call connection.
What general properties should this sound have to work here? Voices sound pretty good on the connection but default notification sounds all tend to clip and distort horribly.
If voices sound good over the conference call connection, then try to make your notification sound have properties that are similar to the human voice. For example, a similar frequency spectrum — typically centered around 1 kilohertz. If you can say “ah” over the connection and it sounds good, then a sound that sounds like that “ah” will likely sound good.
You probably don’t want to make your sound too short, because these systems have noise gates that come in and out. If a sound is too short, it may just get gated out.
Also keep in mind that you don’t necessarily have to normalize your sound and fill all of the gain. Leave some headroom — fill 75% of the gain, and that may solve some or all of your clipping problem. If the sound is clipping, you can try just filling 25% of the gain.
There are many things that could account for the distortion you are describing, and I don't know if it is efficient to try and diagnose this. It could be due to the eq distribution of the content as well as overall volume. Seems better to me to just start over with a simple sine tone and see if that can work.
Are you generating the sounds procedurally or using prerecorded cues (e.g., wav, ogg or mp3)? If making the sounds via a sine function, it is good to remember to put in fade-in and release intervals so the sounds don't make an ugly click when they start or stop. For example, a bell-like envelope with a short attack (3 millis, say) and longer decay (250 to 333, ballpark starting point) could work without sounding unpleasant.
It seems to me a quickly repeating note (e.g., tapping two or three times) pitched around 1 kHz would probably be in the ballpark. This tone would be within speech frequency ranges, at the upper end (for vowel formants, though not as high as the fricatives or sibilants which are often shorted in low-quality audio) where it would grab attention. A little note pattern could also work, e.g., C7, E7, C7.
There's a free little Java tool made to make utility sounds--DinkMaker. If you download it (it is small, I wrote it, requires Java 8), I'd recommend trying the file named DinkA6, and playing around with the settings. The envelope is a little short--I'd lengthen the release.
It is easy to try different frequencies with this starting patch, and find out what different pitches sound like and what their frequencies are. The tool will let you export to wav if you come up with something that you like. A multi-tone cue would either be played procedurally, piece by piece in quick succession, or the parts could be assembled into a single cue with Audacity and imported and played via a single command.
