July 2016


Why Do WAV And FLAC Files Sound Different?
Article By Dr. Charles Zeilig And Jay Clawson
Page 2

  So, at this point, what can be concluded from these results? We can say with some confidence that the metadata-associated cover art is primarily responsible for the decline in height reproduction and associated sound quality when WAV and FLAC files are interconverted. Without a large enough memory allocation in the playback software, this loss of replay sound quality is observed in both FLAC and WAV files. Whatever this property of the metadata might be, remarkably it is transferable from FLAC to WAV and back again to WAV, and with increasing negative effect with the number of conversions. Moreover, this particular file property cannot be detected as a change in file size or by conventional null testing procedures, yet it can be readily detected by ear and may be quantified by observed changes in vertical height reproduction.

We wondered what effect metadata size itself might play in determining the degree of negative influence on WAV or FLAC reproduction quality. Are there any industry standards for metadata size variations in high resolution downloads? A random survey of many such HD audio files accumulated over several years from various suppliers found large size variations in metadata, with no apparent correlation with the music type, whether pop, classical, or anything in between (see Table 1 in Part 2 of the unabridged article).

To examine this question in more detail we created our own sets of España metadata files, starting with the manufacturer's supplied cover art. Two series of images, comprising either different resolutions or different compression levels, were created using Photoshop tools. Each series produced a range of increasing file sizes, which were then attached to our test music track from which all metadata had been previously removed (at least that which could be discerned by using the Properties file in Windows OS). The results of this experiment are seen in Fig 5.

To a first approximation, there's an obvious correlation between total metadata size and observed image height, caused by either increasing pixel resolution or by increasing jpg compression settings. In the Photoshop program used to generate each series, the greater file compression (blue line), relates to the smaller the file size; the greater the pixel density or resolution, the greater the file size and thus the poorer the height reproduction (orange line). The different shapes of each height results curve suggest that there may be two different mechanisms involved in the degradation of FLAC quality. Where the two curves intersect gives the point at which an optimisation of metadata compression and resolution should yield the best sound quality and height reproduction.

We put this prediction to the test by conducting the experiment shown in Fig 6. We created our usual WAV-to-FLAC-to-WAV conversion file series but here show the results in bar graph form, ranging from the initial maximum quality WAV and the first and last converted FLAC files.

We were somewhat surprised by the implications of these results. First, if one looks at the WAV results (solid bars), even with memory playback engaged in JRMC, a statistically significant negative effect related to the size and compression level of the original cover art is observed. This loss is present when it is compared with either no metadata or our optimised home-made metadata. Nevertheless, there is no sensible statistically significant difference between these three files. This result shows that improperly optimised metadata has the potential to reduce even WAV file performance and sound quality to a degree.

Examining the effect of metadata on FLAC files from the beginning (upward hatched bars) to the end of the WAV-to-FLAC-to-WAV conversion sequence (downward hatched bars) under the same playback circumstances, with the original complete metadata supplied by the manufacturer, we see the same declining pattern as shown previously in this article (black hatched bars). Optimising the metadata significantly improves height reproduction (blue hatched bars) but not to the same extent as for both the first and fifth FLAC files.

The greatest improvement in FLAC performance occurred with the removal of the metadata, regardless of its presence or absence (red and green hatched bars). Even with this improvement, no FLAC file reaches statistical equivalence with the WAV file if it contains the original metadata (black bar). These results suggest that there is at least one more factor besides cover art that is involved in stunting the full retrieval of WAV-equivalent performance. Assuming that a recovered WAV file from a previously compressed FLAC file is a bit-perfect replica of the original WAV file, then it seems likely that there is still something in the metadata, which we had no way of accessing, and was interfering with full recovery of audio equality.

Examination of subjective image height measurements for some of the larger FLAC and WAV files (Figs 4 and 5) reveals massive subjective height differences in the range of 50-60ins, and we have struggled to describe the subjective impact.

More intrepid readers could do the following experiment on their own systems: obtain a DXD high resolution download and its matching CD (for example, www.2l.no/hires/ provides WAV format DXD samples recorded at 24-bit/352.4kHz resolution). Using a good quality DAC, compare the sound of the CD replayed on a dedicated disc transport with the sound of the DXD file (played back from a desktop or dedicated computer/server). What you should hear is something that approximates the largest height differences we have observed, and it comes as quite a shock to realise how much sound quality the industry may be losing by using poorly prepared FLAC files. We consider that the sound quality and emotional benefits that may be obtained by optimising music files can be thrilling to say the least.

FLAC
Finally, in a previous article (TAS issue 248) we found that FLAC height reproduction and sound quality was also degraded by the degree of FLAC compression. We repeated this experiment for this report and found similar compression-related losses, regardless of the presence or absence of metadata (see Part 2 of the unabridged article available on the HIFICRITIC website), constituting yet another negative variable affecting FLAC performance.

Conclusions and Discussion
To summarise, we have found at least five different factors that can explain why the lossless FLAC compression format degrades the sound quality of previously uncompressed WAV files (see Box1). These independently configured factors may act together and can result in a disturbingly large loss of sound quality. While the metadata effect and its audible losses can be minimised or reversed, the FLAC compression and decompression effects appear to be permanent unless replaced by clean metadata that has never been exposed to prior conversions. The mechanism for the sound quality damage of all five factors used for FLAC results from their interaction with computer/server replay processes. The effects of this extra loading on computer operations, although apparently small, must act to degrade the computer-processed data then sent to an external DAC. As discussed by others in the field, these negative contributions are thought to involve several forms of jitter, also power supply fluctuations, and RFI (electronic noise interference) which ultimately compromise the optimal functioning of the digital and analogue sections of the DAC.

On the basis of our results thus far, we can make specific recommendations to the music download industry for issuing improved music files:

• Erase metadata containing historical records of previous repeated WAV-to-FLAC-to-WAV conversions and replace with fresh metadata prior to final distribution.

• Since storage capacity is so cheap these days, unless there are bandwidth or download speed issues, or maintenance of metadata concerns, do not use compression (ie, use WAV or AIFF in place of FLAC or ALAC).

• Use a download manager that can add fresh metadata from internet sources just before writing to the customer's storage device.

• If you must use the FLAC format, supply files created using the dBPA software with the uncompressed 'U' setting, or no less than the 0 setting. (Never follow the dBPA recommendation to use setting 5.)

• Make metadata-associated cover art no larger than 800x800 pixels with maximum or near maximum jpg compression (similar to Photoshop QL0 or QL1). (When ripping your own CD collection, be sure to follow this same advice for best sound quality)

Why Lossless FLAC Compression Degrades Uncompressed WAV File Quality
1. Resolution of metadata associated art (MDA).

2. Degree of MDA compression as created in typical photo-editing software.

3. Degree of FLAC compression according to settings in dBPowerAmp.

4. CPU load during decompression and conversion of FLAC files to PCM format.

5. Allocated buffer size in playback software, music server, and/or digital to analog converter.

Readers are encouraged to refer to an expanded discussion of additional factors in Part 2 of the unabridged version of this article which is available for download from the HIFICRITIC website.

---> Back To First Page