What does 25ms sound like?

Building upon my initial tests, I wanted to find a better way to measure the latency of audio devices, and what this actually sounded like. As described in this article, an easy way to do this is by performing a round-trip audio feedback (Larsen effect) test using Audacity.

I used the same PC (Intel Core i5-8500) running Audacity on Windows 10 Pro 2004 to run a few simple tests:

  • Yeti USB microphone feeding to my USB Prime headphone amp
  • Corsair USB headset microphone to same headphone
  • Scarlett 2i2 (2nd gen) input to output (this was the only one I could test using a loopback cable)

I used Audacity to create a simple mono rhythm track for each test. Below are the results, along with WAV audio files combining the generated track (left channel) with recorded track (right channel). You should be able to hear an audible “echo” in the right channel as a result of the delay.

Yeti + Prime (31ms)
Corsair Headset (28ms)
Scarlett 2i2 (13ms)

In case you may wonder what it sounds like, here is the same rhythm track with the right channel shifted 25ms:

25ms delay on right channel

These all used MME at 44Khz, 32bit. I tried DirectSound and WASAPI but for some reason these increased latency times for my Yeti+Prime test to about 100ms and 500ms, respectively.

It surprised me that the Corsair Headset measured a lower latency than my Yeti + Prime combo, contradicting my previous test’s observations. This is a different PC than my previous test so I wonder if there is something different with the Kernel (this one is Slow Ring versus previous was one runs the last public release), settings, drivers, or just the software I was using. Or, if it’s just my own bias coming into play. I suppose this emphasizes the need for more measurements and controlled tests versus just listening to how it sounds.

Perhaps the disappointing result here is that the round-trip latency measurements for the commercial gear is already too high: at or slightly above our overall goal. Of course, the type of music is material. It’s probably fine for slow tunes but challenging for songs with lots of staccato. But this is all on the same PC. It leaves no extra room at all for network latency, which we know plays a significant factor.

One conclusion is that our success likely depends on the use of pro audio gear, like the Scarlett 2i2, to leave enough leeway for the added networking latency. Considering the solo version of this costs about $110, it sets a relatively high entry point. Perhaps this is not a challenge for professional musicians, but it may be a larger hurdle for kids in childrens choirs.

For comparison, HiFiBerry’s DAC+ ADC board is about $50 and claims online suggest single-digit round-trip latency. Using Elk Audio OS (free), you can supposedly get it as low as 1ms. For under $100, you could build a complete box that leaves a lot of room for network! I have a few of these on order to test out, which hopefully will arrive this week. I look forward to running similar tests and posting those results soon.

One thought on “What does 25ms sound like?

  1. Update: I followed these instructions for testing with Audacity, but realized today that I skipped the “set latency compensation to 0” step. My latency compensation was sent to the default of -130, which seems to mean these numbers (and all the sound files) have a gap too small by 130ms.

    The revised measurements would be:

    * Yeti+Prime: 161ms
    * Corsair: 158ms
    * Scarlett 2i2: 143ms

    I also realized that I was using the default drivers versus Focusrite’s driver for the Scarlett 2i2. After installing that, the latency got worse! With drivers, my MME measurement (100ms buffer) was 184ms and my WASAPI measurement is 174ms. After lowering my latency buffer from the default of 100ms to 50ms, I was able to get the same 174ms measurement for WASAPI. All are extremely disappointing results. It also has me questioning the validity of using Audacity, my Windows PC, etc.

Leave a Reply