A better Dolby Pro Logic II downmix

bleo · 21st July 2003, 14:16

The Dolby Pro Logic II downmix used in programs such as Azid.dll, BeSweet, HeadAc3he and AC3Filter, was derived in this thread: http://forum.doom9.org/showthread.php?s=&threadid=27936

A test sample of 5 channels active one at a time, downmixed to 2 channels with this matrix, should be able to be decoded by a Dolby Pro Logic II decoder back to the original 5 channels. However, I have been disappointed by the rear channel separation of the current DPL2 downmix.

The rear channel separation can be improved by altering the 'balance' of the surround channel downmix while maintaining constant power. That is, BL is downmixed more to Lt, while BR is downmixed more to Rt. However, by skewing the balance too much to either side, we lose phase information, resulting in reduced front-back separation.

Here I describe a method for deriving a better DPL2 downmix

Methods
Contents of 5.1 AC3 test source 'Avia ac3test' (http://www.digital-digest.com/dvd/do.../trailers.html)

Code:

time (sec)	channel		contents (announcer says: )
12-14		L		'left front'
14-16		C		'centre'
16-18		R		'right front'
18-20		BR		'right surround'
20-22		BL		'left surround'
22-24		L		'LFE'
		LFE		[rumble]

'surround2' downmix
"C:\tmp\BeSweetv1.5b19\BeSweet.exe" -core( -input "c:\tmp\ac3test.DELAY 0ms.ac3" -output "c:\tmp\ac3test.surround2.wav" -2ch -logfilea "C:\tmp\BeSweetv1.5b19\BeSweet.log" ) -azid( -s surround2 -L -3db )

'4.8 dB' downmix
Using AC3Filter in GraphEdit. Exact values of surround downmix coefficients are SQRT(3/4) and SQRT(1/4). The surround power difference is triple, hence 4.8 dB

Lt = L + 0.7071 C + 0.7071 LFE - 0.866 BL - 0.5 BR
Rt = R + 0.7071 C + 0.7071 LFE + 0.5 BL + 0.866 BR

'6 dB' downmix
Lt = L + 0.7071 C + 0.7071 LFE - 0.8944 BL - 0.4472 BR
Rt = R + 0.7071 C + 0.7071 LFE + 0.4472 BL + 0.8944 BR

DPL2 decoding as per thread: http://forum.doom9.org/showthread.php?s=&threadid=57736

Notation of decoded channels: L' R' C' LFE' BL' BR'

Screenshots taken from Cool Edit Pro

Separation between channels calculated from difference in total RMS power of 2 sec segments

Results
Figures at http://www.geocities.com/bleo98

Figure 1: surround2 BL'
18-20 sec, BR leaks into BL'. Separation ~ 9 dB

Figure 2: 4.8 dB BL'
18-20 sec, more separation between BL' and BR' ~ 26 dB

Figure 3: 6 dB BL'
18-20 sec, even more separation between BL' and BR' ~ 31 dB

Figure 4: 6 dB L'
BUT! 20-22 sec, BL leaks into L'!

Figure 5: 4.8 dB L'
20-22 sec, no leakage of BL into L'

Discussion
The 4.8 dB downmix provides optimal separation of all 5 channels. I now use this matrix in AC3Filter, but should the azid.dll matrix be changed?

surround2 is quieter because the downmix coefficients were normalised to prevent arithmetic overload. The 4.8 dB and 6 dB coefficients are not normalised. Overload would only occur if all 6 channels were at full scale during EXACTLY the same time sample, which is highly unlikely. In fact, this did not occur during testing with the Lord of the Rings soundtrack.

The 'pulses' every two seconds are steering artifacts: a delay in steering inherent in the decoder's feedback design (?)

DPL2 cannot steer 5 independent channels simultaneously. For example, during 22-24 sec, L contains 'LFE' and LFE contains [rumble]. Note, LFE is supposed to be decoded to C'. However, DPL2 steers BOTH to L' AND C'.

21st July 2003, 14:16	#1 \| Link
bleo Registered User Join Date: Jul 2003 Posts: 78	A better Dolby Pro Logic II downmix The Dolby Pro Logic II downmix used in programs such as Azid.dll, BeSweet, HeadAc3he and AC3Filter, was derived in this thread: http://forum.doom9.org/showthread.php?s=&threadid=27936 A test sample of 5 channels active one at a time, downmixed to 2 channels with this matrix, should be able to be decoded by a Dolby Pro Logic II decoder back to the original 5 channels. However, I have been disappointed by the rear channel separation of the current DPL2 downmix. The rear channel separation can be improved by altering the 'balance' of the surround channel downmix while maintaining constant power. That is, BL is downmixed more to Lt, while BR is downmixed more to Rt. However, by skewing the balance too much to either side, we lose phase information, resulting in reduced front-back separation. Here I describe a method for deriving a better DPL2 downmix Methods Contents of 5.1 AC3 test source 'Avia ac3test' (http://www.digital-digest.com/dvd/do.../trailers.html) Code: time (sec) channel contents (announcer says: ) 12-14 L 'left front' 14-16 C 'centre' 16-18 R 'right front' 18-20 BR 'right surround' 20-22 BL 'left surround' 22-24 L 'LFE' LFE [rumble] 'surround2' downmix "C:\tmp\BeSweetv1.5b19\BeSweet.exe" -core( -input "c:\tmp\ac3test.DELAY 0ms.ac3" -output "c:\tmp\ac3test.surround2.wav" -2ch -logfilea "C:\tmp\BeSweetv1.5b19\BeSweet.log" ) -azid( -s surround2 -L -3db ) '4.8 dB' downmix Using AC3Filter in GraphEdit. Exact values of surround downmix coefficients are SQRT(3/4) and SQRT(1/4). The surround power difference is triple, hence 4.8 dB Lt = L + 0.7071 C + 0.7071 LFE - 0.866 BL - 0.5 BR Rt = R + 0.7071 C + 0.7071 LFE + 0.5 BL + 0.866 BR '6 dB' downmix Lt = L + 0.7071 C + 0.7071 LFE - 0.8944 BL - 0.4472 BR Rt = R + 0.7071 C + 0.7071 LFE + 0.4472 BL + 0.8944 BR DPL2 decoding as per thread: http://forum.doom9.org/showthread.php?s=&threadid=57736 Notation of decoded channels: L' R' C' LFE' BL' BR' Screenshots taken from Cool Edit Pro Separation between channels calculated from difference in total RMS power of 2 sec segments Results Figures at http://www.geocities.com/bleo98 Figure 1: surround2 BL' 18-20 sec, BR leaks into BL'. Separation ~ 9 dB Figure 2: 4.8 dB BL' 18-20 sec, more separation between BL' and BR' ~ 26 dB Figure 3: 6 dB BL' 18-20 sec, even more separation between BL' and BR' ~ 31 dB Figure 4: 6 dB L' BUT! 20-22 sec, BL leaks into L'! Figure 5: 4.8 dB L' 20-22 sec, no leakage of BL into L' Discussion The 4.8 dB downmix provides optimal separation of all 5 channels. I now use this matrix in AC3Filter, but should the azid.dll matrix be changed? surround2 is quieter because the downmix coefficients were normalised to prevent arithmetic overload. The 4.8 dB and 6 dB coefficients are not normalised. Overload would only occur if all 6 channels were at full scale during EXACTLY the same time sample, which is highly unlikely. In fact, this did not occur during testing with the Lord of the Rings soundtrack. The 'pulses' every two seconds are steering artifacts: a delay in steering inherent in the decoder's feedback design (?) DPL2 cannot steer 5 independent channels simultaneously. For example, during 22-24 sec, L contains 'LFE' and LFE contains [rumble]. Note, LFE is supposed to be decoded to C'. However, DPL2 steers BOTH to L' AND C'. Last edited by bleo; 27th July 2003 at 11:38.