from BlackBird.wav
(bitrate = 128 kbit/s)
Sound Forge 4.5h,
highest quality
(bitrate = 128 kbit/s)
Notes by ff123
On the Lame Quality and Listening Test Information page, there is a test sample called "BlackBird.wav," which has the following description:
From davel at caffeine.co.uk: The opening bars from the Beautiful South track 'Blackbird on the Wire' - you will notice it has some very sudden synthesized drums in it. I encoded this track using both Lame and Frau, and found that Frau just couldn't handle the first drum hit (it even had trouble with subsequent ones which are less apparent on this particular example) and actually drops out for a fraction of a second before the drum hit. I have tried many settings on both decoders and Lame seems to handle it no matter what, whereas Frau always trips up, even with the highest quality setting. Bitrate is immaterial.
I have run across other samples of music in which dropouts are apparent. For example, see my listening test of a Radiohead sample, at 128 kbit/s. The dropouts don't necessarily have to sound like momentary losses of sound; sometimes they manifest themselves as chirps or "knocking" sounds. The dropouts do get better with increasing bitrate, contrary to what is implied above. FhG's Alternate codec produces similar dropout artifacts in the same samples, whereas no other mp3 codec I have tested produces such dropouts. This leads me to believe that this is some sort of bug in FhG's reference codec, and that MP3Enc and Alternate share the same bug because they share the same codec heritage.
On most musical material I've encoded with MP3Enc at 192 kbit/s, I do not hear any dropout artifacts at all. However, when I recently played with Elton John's Blue Eyes, I heard a dozen such artifacts throughout the song at that bitrate. Using the spectral view mode of Cool Edit Pro, I was able to pinpoint the location of the most obvious blips and to show what they look like in graph form. I could then check for potential problem areas in the rest of the file. Some of the potential dropouts I located this way were audible, others were masked by the music. The following graph shows what the dropout looks like in the BlackBird.wav clip for MP3Enc31 at 128 kbit/s and for FhG's Alternate codec within Sound Forge 4.5h at 128 kbit/s (click on thumbnail images to enlarge).
The dropout in BlackBird.wav is seen below as the narrow, dark area within the transient (a hi-hat), lasting for about 1 to 2 milliseconds and extending up to about 6 kHz. Dropout artifacts are invariably associated with music transients (sudden bursts of energy across the whole frequency spectrum).
| mp3 Encoder | Spectral View of selection from BlackBird.wav |
| MP3Enc31 -qual 9 (bitrate = 128 kbit/s) |
|
| Alternate Codec: Sound Forge 4.5h, highest quality (bitrate = 128 kbit/s) |
|
In Elton John's Blue Eyes, which is 3:30 minutes long, I located, and verified that I could hear, 12 locations within the mp3 file (encoded with MP3Enc31 at 192 kbit/s) which had a spectral plot similar to the one shown above. In the plot below, I show one such location. In this one, the artifact is not a dropout, in the sense that I can hear the music pop out and back in, but rather the artifact is a distinct sound, like someone knocking on wood. You can find this clip, BlueEyesTrimmed.mp3, on my Audio Samples Page. The artifact shown below occurs about 1.85 seconds into the clip. Interestingly, I had to use mp3Trim to slice out the relevant portion of the encoded song -- encoding the equivalent .wav excerpt fails to produce an audible blip. You can get the equivalent .wav excerpt of the original if you'd like to compare it against the mp3 clip. It's on my Audio Samples Page as BlueEyesExcerpt.flac. It has been compressed using the lossless compressor, FLAC. (note: the editing I performed was just an eyeball, and isn't exactly equivalent to the mp3 file).
| mp3 Encoder | Spectral View of selection from Elton John's Blue Eyes |
| MP3Enc31 -qual 9 (bitrate = 192 kbit/s) |
 |
This particular song seems to be unusual in that the dropout artifacts are present and audible at 192 kbit/s. I went through many of the other files I have which were encoded with MP3Enc at 192 kbit/s, and wasn't able to see or hear anything nearly as bad as this song. However, at 128 kbit/s, this problem is prevalent enough that I would hesitate to use MP3Enc at this bitrate, even though many people with much better hearing than me would choose MP3Enc as the best-sounding mp3 encoder at 128 kbit/s.
Addendum (2-14-01): Another problem song at 192 kbit/s: Steely Dan's Time Out of Mind. There are probably lots more.
On an unrelated note, Elton John's Blue Eyes is just about the most horrible-sounding piece of music I've run across. It's got all sorts of extraneous sounds in it which sound like encoding artifacts to me, but which are really there in the original. I don't know what those sounds are, but they're irritating as hell.
Summary
In certain musical selections, the percentage of which may vary depending on your musical tastes, the Fraunhofer Mp3Enc/Alternate codecs have a bug/quirk which introduce short, low-frequency losses of energy into the encoded output. These may or may not be audible, depending on the music. The artifacts may manifest themselves as momentary dropouts or as chirping/knocking. The artifacts tend to become less noticeable with increasing bitrate. It is possible to search through a file using the spectral view mode of your sound editor to locate possible problem areas.