Sound Device's 24 bit vs 16 bit Voice Recording Test    
Tim Nielsen directed NatureRecordist List readers to an interesting study posted by Sound Devices titled, "Real World Advantages of 24-Bit Recording." Their test was made with a Shure KSM44 microphone and a SD 744T recorder using studio voice recordings producing these characteristics and post-recording modifications:    
click on waveforms to play soundfile
(1) 24 bit recording with peak saturation at -13.8dB:
(2) 24 bit recording with mic pre gain reduced 40 dB and digitally normalized for a matching -13.8dB peak:
(3) 16 bit recording with peak saturation at -13.7dB:
(4) 16 bit recording with mic pre gain reduced 40 dB and digitally normalized for a matching -13.7dB peak:
(5) 16 bit low-scale sound file intercut with the low-scale 24 bit file to make differences easier to assess:
[alternative -1.5mb QuickTime .mov]
Sound Devices concludes, "... the tracks recorded 40 dB down are very different. Most important is how usable the low level 24 bit signal is after normalization. The [-40 dB 24 bit recording] is definitely noisier than the full gain recording but it held up quite well. The [-40dB] 16 bit recording highlights how important recording full scale digital is in a 16 bit environment."    
The test prompted me to ask: Is it appropriate to apply these findings to circumstances where low-scale recordings are typical rather than the result of under-recording as in the SD test? Would 16 bit recordings exhibit more quantizing noise than 24 bit recordings when the mic pre gain is set near maximum in very quiet settings?    
  Note: I did not incorporate SD's tests in this report which used dithering during recording. Dithering tends to improve 16 bit performance so my test examines the noiser of the samples.    
Remote Location Nocturnal Ambience Test    
The place where I made the following nocturnal recordings in rural Wisconsin can become quiet enough in the midde of the night to capture subtle sounds coming from considerable distance when using low-noise equipment. The quietest, background passages produced peak levels of -40dB measured in sample editor of Logic 7.2. The sound maintaining the background peaks seems to be an idling engine about 1/2 mile away. The low-amplitude effects are water drips from foliage a half hour after light rain and a Barred owl calling over a ridge that is an eighth of a mile away. The mics used were Rode NT1-A's with 6dBA self-noise and 25mV/Pa sensitivity. The maximum record gain of 70dB was used on the SD 744T recorder.    
click on waveforms to play soundfile
(6) Original 24 bit/48K field recording: NT1-A's->744T at maximum gain. Peak of quietest background at -40.4dB
(7) Original 16 bit/48K field recording: NT1-A's->744T at maximum gain. Peak of quietest background at -40.1dB
(8) 16 bit and 24 bit segments are alternated starting with the 16 bit sample. Edits occur after each owl call. Gain boosted 20dB to aid assessment:
[alternative - 3.1mb, 24bit/48K .wav]
Practical Sound Levels and Quantization Noise    
It was not evident to me at first, but SD's test method of lowering the mic pre gain 40 dB to produce the low scale files might not be consistent with sound levels recordists will encounter when mic pre levels are appropriately set. In studying their recordings, I noted that room tone in SD's voice recordings, where the quantizing noise is most evident, is about 25dB softer than the voice peaks-- residing at about-39dB.    
To get their low scale examples, SD reduced the 744T's mic pre level 40dB which should have placed the peaks in the studio room tone in the neighborhood of -79dB. In contrast, the peaks in the softest background sounds in my nocturnal ambience recording are about 40dB higher.    
Assuming the background sound levels in the rural, night setting were around 25dB, applying a 40dB gain reduction to create the low-scale sound samples like SD's would require sound levels around or below the threshold of human hearing. Quantizing noise does exhibit considerable treble emphasis which would make the noise more audible against sound sources with low-frequency weighting.    
At what sound level does 16 bit Quantization Noise become an issue?    
Out of curiosity, to roughly estimate the quantization noise that one might encounter when recording sounds of absolute lowest audibility with 70dB of preamp gain and the NT1-A's, I digitally reduced the gain of my 16 and 24 bit nocturnal recordings by 20dB:    
  This resulted in the expected --60dB peaks in the background sounds in the rural nocturnal recordings. I then boosted these low-scale recordings 40dB to match the playback level of sample (8). The below sample approximates how the 16 and 24 bit formats might respond with the 16 bit recording first and changing after each owl call:    
click on waveform to play mp3 version
(9) 20dB gain-reduced16 bit and 24 bit segments are alternated starting with the 16 bit sample. Edits occur after each owl call.
[ alternative 3.8mb, 24bit/48K .wav]
  To explore a hypothetical* sound level where quantizing noise in the 16 bit file would become more audible, I digitally reduced the original -30dB which resulted in the expected -70dB peaks in the background sounds. These very low-scale recordings were digitally amplified 50dB to match the playback level of samples (8) &(9):    
(10) 30dB gain-reduced16 bit and 24 bit segments are alternated starting with the 16 bit sample. Edits occur after each owl call
[ alternative 3.1mb, 24bit/48K .wav]
* Due both to sound levels falling below 0dB and technical limitations of microphone self-noise and noise bed performance of mic preamplifiers.
  (A) I had predicted that 16 bit quantization noise might be audible when recording in the quietest natural locations and I was wrong. It remains surprising to me that quantization noise is inaudible with the 16 bit file all the way down to levels producing -60dB peaks (or ~.1% saturation!) from the background sounds. One would think, as the bit "steps" are divided evenly across the 96dB total range, that resolution and performance would drop off faster than this.    
  (B) It still makes sense to me to record with higher resolution if one's equipment and budget permits it, but these results suggest to me that 16 bit recorders with low-noise and reasonably high-gain preamplifers should be able to avoid quantization noise under real world conditions. If recording media is scarce or long takes are required, 16 bit recording might be a very good, essentially, uncompromised solution.    
  (C) If one owns a 24 bit recorder with mic preamp gain under 45 dB, creating 24 bit files in the field might help avoid quantizing noise when collecting material in quiet locations.    
  (D) I often run 16 bit and 24 bit rigs at the same time and many times with the same model mics. I cannot consistently distinguish 16 bit from 24 bit originals of ambience recordings made in quiet locations using headphones or speakers. It was my hope that 24 bit resolution would improve clarity between 125Hz and 700Hz (for better spatial imaging), but mic performance in this range appears to be the greater quality limiting factor than bit depth.    
  (E) When I'm mixing and using equalization, the 24 bit files are more responsive than 16 bit files and the 24 bit original seems to "hold" or preserve more of what I hear in Preview mode after the mix file is digitally-rendered. Whether there is a significant difference between a 24 bit field original and a 16 bit file up-sampled to 24 bits before processing, I haven't tested.    
Low-Scale Digital Audio Recording
in Quiet Locations
by Rob Danielson 2007.07
Preliminary: This study uses comparative listening tests to examine only the role of amplitude quantization noise in the recording stage and not other possible advantages of 24 bit files. Only the analog to digital processing achieved with a Sound Device 744T recorder is examined. All of hre measurements below were made in the Logic 7.2 sample editor with a 96dBFS reference.