Review of the Korg MR-2

The Korg MR-2 does not live up to its premium/professional price.  When a field recorder costs as much as a smartphone, you expect it to be reliable and high performance.  I’ll list the flaws first, state my recommendations, and then delve into the testing details.  Each of these flaws except the last two have caused me to lose performances that I attempted to record.

An illustration of how the MR-2 fails is the complexity of the instructions I had to write for this show:  Apparently the instructions for the decibel level were too complex:

Luckily, I had correctly guessed the decibel level in advance and the nice couple who recorded the show for me did everything else perfectly.

Also, the main takeaway from this post, applicable to all audio equipment, is that 24-bit PCM recording does not mean you can amplify by 48 dB and still have clean 16-bit PCM.  The noise coming from the mic, amps, and ADC gives 16 bits (96 dB) at best to start with.  This guy suggests that professionals record at 24 bits, but it must only be in a studio environment with gear that costs thousands or tens of thousands of dollars: When does 24 bit matter?

List of Flaws

  1. When the input level is set below -18 dB, the clipping indicators fail to function.  Setting the input level below -18 dB never makes any sense, so the software should not allow it.
  2. The recorder doesn’t actually function on 2 AA cells.  With fully charged NiMH cells, there isn’t enough voltage to run the Korg.  With alkaline cells, it runs for a short while, but alkaline cells discharge with a linear drop in voltage and the Korg shuts off while they still have well over 50% charge left.  I ended up buying 1.6 V NiZn cells to make the Korg work.
  3. The USB jack is a mini USB, not micro USB, so it won’t work with modern phone cords.
  4. The Korg freezes up unpredictably.  When it does that, you have to remove the batteries to restart it.  Try doing that while it’s screwed into a stand and you’re scrambling to restart so you don’t miss too much audio.  It’s almost better to run it without any batteries so you can just pull out the USB power, but how is it acceptable for a recorder to freeze up at all in the first place?
  5. When a WAV file is initialized, it writes a length of 0.  Once it runs into the 1 GB limit, the Korg writes the proper size of the file and starts a new file with the same size as the 1 GB.  This means that if it freezes during the first 1 GB, it takes technical skills to repair the 0 length of the WAV file, but if it freezes in the next gig, everything works fine.
  6. When the red “record” LED is blinking, the recorder doesn’t actually record.  You have to press the “record” button again to make the LED stay on steadily and record.
  7. The term “high resolution mobile recorder” is incorrect.  24-bit resolution is misleading because there’s no way to get that quality of audio to the ADC in this thing.  24 bits implies 144 dB of dynamic range.  Or that you should be able to get an 8-bit quality recording at a level that’s 96 dB quieter than the peak volume.  In testing, the Korg demonstrated absolutely no way to get more than 15 bits worth of information recorded and that’s being generous.  If you record at 24 bits instead of 16 bits, the least significant 8 bits will be completely useless.  You can write anything you want to them, since they’re drowned out by noise in the electronics.


  • Korg can solve issue 1 by modifying the firmware to disallow input levels below -18 dB.  They can also update the user manual to explain what the input level adjustment and the mic sensitivity setting (Low, Med, High) do.
  • I don’t know if there’s a way to fix issue 2, that the recorded shuts down at too high a voltage.  It may require a recall or a warning to run the Korg off USB power.
  • Issue 4 probably is due to an error in the firmware.  There must be a bug somewhere, but I’m sure it’s expensive to try to figure out where.  The engineers need to go back to testing the device in debug mode to figure out what’s happening.
  • Issue 5 can be fixed by having the maximum possible recording length written to the WAV file and then modified to the right amount once the recording stops.  That way, if the power gets cut or the recorder freezes in the middle of a recording, the audio is still easily read.
  • Issue 6 can be fixed by changing the standby blink pattern to something other than a digital SLR’s record blink.  The LED could flash rapidly three times in a row, with pauses in between, sort of as a warning that it’s not yet recording.
  • And a workaround for issue 7 could be to create a custom audio format that uses AGC (Automatic Gain Control) and records the upper 16 bits, with the lower 8 bits used to encode the gain level used.  Then some custom software on the processing side could lower the volume of the quieter (high gain) sections to match relative to the louder (low gain) sections and reconstruct a 24-bit recording.  That might be tough, though, because gain settings don’t change instantaneously and the transitions could create audible clicks.  Those transitions could be repaired in software, though, using whatever technique Audacity’s “Repair” effect uses.  Update: It turns out that the idea of recording simultaneously with 2 ADC’s and recording extra level data to a file already exists:

MR-2 Teardown

Opening up the MR-2, you find a $6 Burr Brown PCM1804 chip at the heart of things.  Reading the datasheet for the PCM1804 reveals that the chip has some fancy features, such as DSD and high data rate PCM capability.  The SNR is a decent 111 dB.  Except that number only applies when it receives a full scale signal, with the noise remaining constant as the signal level drops.  It’s apparently not possible to set lower reference voltages for the ADC to record lower amplitude signals and expect the noise to fall accordingly.

IMG_5107 IMG_5110 IMG_5104

 Test Procedure

I began with a recording of Sail Away Ladies, performed by Brittany Haas.  I also wrote some Python code that could halve or double the values stored in a WAV file.  I used it to create a version of the audio sample that got quieter by 6 dB (1 bit) every 2 seconds.  My test equipment consist of a Woo Audio WA-7 Fireflies headphone amp hooked up to Sennheiser HD-650 headphones.  I pipe the output from the tube amp to the Korg through a Monster 1/8″ stereo cable because cheap unshielded cables pick up noise, at least in my car.

Test 1 – Noise becomes audible at 42 dB SNR.

For the first test, I took the sample that loses 1 bit every 2 seconds and amplified it back by 6dB every 2 seconds.  This effectively clears out the lower bits of the audio, reducing the quality one bit at a time.  This is what the resulting waveform looks like.  You can visibly see the discretization effects of up to 4 bits.

cutting bits every 2 seconds

After 34 seconds, when 17 bits have been cut and 7 are remaining, I definitely hear a little noise in the background, sort of like the hiss we used to hear on cassette tapes.  So I’d say that 8 bits or 48 dB of SNR is sufficient for the noise to be inaudible to me.  Of course, this only applies for a constant signal level, so wouldn’t work for a classical recording with loud and quiet sections.  My results make sense, given that vinyl and cassette tapes have about 50 – 55 dB of SNR.  They certainly sound fine during sections of full loudness, but the recording medium noise becomes apparent in any quiet sections.

 Test 2 – Compared to Low mic sensitivity, Medium is about 8 dB and High is about 14 dB louder, the Line input is 30 dB quieter.

I set the audio output level of the WA-7 such that it causes the Korg’s level meters to read around -3 dB on the Low mic setting with 0 dB input level.  Then, I switched the mic setting to Medium and then High, adjusting the input level to achieve the same -3 dB peaks.  That required 8 dB and 14 dB settings.  With the same signal going into the Line input, the peaks ended up around – 33 dB, suggesting that Line is 30 dB less sensitive.  For the subsequent noise tests, we’ll want to attempt to record a very quiet and clean signal.  We can use these sensitivity numbers to calibrate the level of the quiet signal to infer how much louder it could get without clipping.

Test 3 – Input Level settings below -18 dB result in unexpected clipping

For this test, I recorded the same loud audio with High mic sensitivity and input levels of -17 dB and – 24 dB.  You can see that the clipping is normal at -17 dB, but occurs at about half the maximum value of the ADC when the input level is at -24 dB.  Basically, if the Peak indicators on the Korg never go red no matter how loud the sound is, you know they’re not working because the mic itself is probably clipping.  Lowering the mic sensitivity and turning up the input level helps avoid this, since the weird clipping always happens around input level settings of -18 dB, regardless of mic sensitivity.

Test 4 – Always use Low mic sensitivity

For test 4, I set the audio level to be relatively quiet, so that it peaks around -10 dB on the High 0 setting.  I recorded the audio sample at High 0, High -14, Low 0, and Low -18 settings.  The corresponding peaks would be at -10 dB, – 24 dB, -24 dB, and -42 dB at Low -18.  With 4 bits of headroom for High -14 and Low 0, 7 bits for Low -18, we estimate the level of noise at the various settings by amplifying them back to full scale and listening for when the noise becomes apparent in the audio samples.  We can also compare whether it’s better to record audio at High -14 or Low 0.

The High 0 recording definitely starts to hum after losing 4 bits.  That’s the most sensitive setting, and this shows that it only has less than 34 dB of usable dynamic range.  That is, when audio gets 34 dB quieter from the clipping level, the recording begins to sound like a 7 bit recording.  Or phrased differently, you only get 13 usable bits at the High 0 setting.

The High -14 recording also starts to hum after losing 4 bits.  This suggests that the input level adjustment is doing its job and the noise must be coming from the mic amp.  But now, there is 48 dB of usable dynamic range, since the recording starts out 24 dB quieter than clipping and becomes noisy another 24 dB later.  That implies about 15 usable bits.

With the Low 0 setting, we get almost the same results!  So lowering the mic sensitivity while simultaneously raising input level doesn’t help reduce noise, unfortunately.

At the Low -18 setting, there’s more of a hiss than a hum, and it comes after losing just 2 bits.

So far, we’ve learned that trying to record quiet sounds with the High 0 setting is pointless.  That using High -14 can capture the same quiet detail while keeping 2 bits more of headroom.  But also, Low 0 is the same as High -14, so really there’s no point in using a mic sensitivity other than Low.  With the sensitivity set at Low, going from an input level of -18 to 0 does capture 2 bits more of detail, so it’s worth adjusting the input level to try to match the sound.

Test 5 – Line input offers 16 bit / 96 dB SNR, but requires an amplified input.

Here, I set the audio input to -3 dB from peak at Low 0.  Then, recorded the same audio with the Line input at 0 dB input level.

As expected, the Low 0 recording exhibits noise after losing 7 bits or 42 dB.  That gives 45 dB of usable dynamic range, in line with the 48 dB we found in the previous test.  This verifies that digitally lowering the audio volume instead of turning the knob isn’t somehow causing more noise to come out of the Woo Audio amp.

The recording at Line 0 dB, after being amplified by 30 dB to bring it back to full scale, became noisy after losing 4 bits.  That means the Line input is least noisy, with 54 dB of usable dynamic range and an estimated 16 bits of SNR.

Final Thoughts

I tried setting my Rode Stereo VideoMic Pro to +20 dB and sending it to the Line input.  I spoke into it with an indoor voice, amplified it by 36 dB, and the quality was decent, with a bit of a hissing noise.  So for recording loud concerts, it seems alright to use the Korg’s line input and turn up the gain on the mic.

The redeeming factor of the Korg is its small, discreet form factor.  If you set the levels right, ensure it has external power, and keep checking that it hasn’t frozen up, it produces good recordings.  The frequency response is flat from 20 Hz – 20 kHz, demonstrated in my next post, so it’s better to use the Korg than on-camera mic input.

Future Testing

Next, I need to test the onboard ADC of my Canon 5D-MkIII in a same way, to see if it performs better or worse.

Then, I want to evaluate the Zoom H6, which claims to offer a -12 dB simultaneous backup recording.  Maybe this is like HDR on a digital camera.  It makes sense to split the input audio and send it to two separate recording circuits that are set to different levels.  The higher gain track can be lowered by two bits in post production and the clipped parts can be filled in with the lower gain track.  I wonder whether the clock timing lines up perfectly, though.  I have some doubts that there’s any point to the backup recording on the Zoom, but it would be great to find out!

I’d also like to evaluate the Tascam DR-22WL, the Tascam DR-70D, and maybe the Nagra SD.  The Nagra looks about as bad as the Korg, but there’s no way to find out because they don’t post any SNR specs or details on the ADC.

It would be cool to have a DR-22WL as something small that you can inconspicuously leave near a stage and control remotely.  And a Zoom H6 or Tascam DR-70D as a multi-input camera-mounted workhorse.  There are essentially three recording scenarios we encounter: set up loads of gear because the venue and band don’t mind, keep the recording equipment small and dark so that you don’t draw too much attention to yourself, and completely discreet recording where the mics peek out of your pocket or are wired over your ears.

Once I find a decent recorder, the real issue will be optimizing the mic.  Upgrading from the Stereo VideoMic Pro to the X would be cool, but probably unnecessary.  I’ll get much more improvement in audio quality by running some white noise the Stereo VideoMic Pro with the dead kitten installed to figure out how to equalize out the muffling of high frequencies.

Leave a Reply

Your email address will not be published. Required fields are marked *