gallaghe@husc8.HARVARD.EDU (Paul Gallagher) (07/14/88)
Here's a question maybe someone out there can answer: Why isn't it possible to completely restore a recording: for example, to remove all extraneous noise (hiss, clicks, coughs), even to make a reasonable guess about information not in the original recording (for example, given a score and a knowledge of the harmonics of a voice or an instrument, to recreate something close to the sound of the original performance)? PG
phd@speech1.cs.cmu.edu (Paul Dietz) (07/14/88)
In article <4944@husc6.harvard.edu> gallaghe@husc8.UUCP (Paul Gallagher) writes: >Here's a question maybe someone out there can answer: >Why isn't it possible to completely restore a recording: for example, to remove >all extraneous noise (hiss, clicks, coughs), even to make a reasonable guess >about information not in the original recording (for example, given a score >and a knowledge of the harmonics of a voice or an instrument, to recreate >something close to the sound of the original performance)? Actually, this sort of thing is commonly done. I remember hearing a story about some Caruso recordings that were restored by having a singer imitate as closely as possible the original, and then using this to generate optimal filters that were then applied to the original recordings. I suggest you look at the literature on adaptive filtering. Widrow and Stearns "Adaptive Signal Processing" in the Prentice-Hall Signal Processing Series is a good place to start. Paul H. Dietz ____ ____ Dept. of Electrical and Computer Engineering / oo \ <_<\\\ Carnegie Mellon University /| \/ |\ \\ \\ -------------------------------------------- | | ( ) | | | ||\\ "If God had meant for penguins to fly, -->--<-- / / |\\\ / he would have given them wings." _________^__^_________/ / / \\\\-
stewarte@sco.COM (Stewart Evans) (07/16/88)
In article <2266@pt.cs.cmu.edu> phd@speech1.cs.cmu.edu (Paul Dietz) writes: >Actually, this sort of thing is commonly done. I remember hearing a story >about some Caruso recordings that were restored by having a singer >imitate as closely as possible the original, and then using this to >generate optimal filters that were then applied to the original recordings. I have heard, admittedly second-hand, of another restoration project (or maybe the same one) which involved computer-modeling of the cone that Caruso sang into (which was apparently much like the horns used as "speakers"), and of its effects on the sound, then applying this transformation in reverse. Needless to say this was a research project, not a commercial endeavor... Anyone know any more about this research? -- "Stupidity is like hydrogen - it's a basic building block of the universe." -- Frank Zappa
max@trinity.uucp (Max Hauser) (07/16/88)
In article <2266@pt.cs.cmu.edu>, Paul Dietz wrote: | In article <4944@husc6.harvard.edu> ... (Paul Gallagher) writes: | >Why isn't it possible to completely restore a recording: for example, to remove | >all extraneous noise (hiss, clicks, coughs), even to make a reasonable guess | >about information not in the original recording (for example, given a score | >and a knowledge of the harmonics of a voice or an instrument, to recreate | >something close to the sound of the original performance)? | | Actually, this sort of thing is commonly done. I'm not sure that I would concur with this. People frequently suggest strategies similar to Paul Gallagher's, above; but the problem arises in translating ideas like "knowledge of harmonics of a voice or an instrument" into hard specifics, algorithms that act on the recorded information, to do the job. The obstacles are not in the broad concept but in the nitty gritty. | I remember hearing a story | about some Caruso recordings that were restored by having a singer | imitate as closely as possible the original, and then using this to | generate optimal filters that were then applied to the original recordings. Stockham at Utah, pioneering the use of serious DSP in digital audio, used a blind deconvolution algorithm fifteen years ago to separate Caruso's "original" voice from the severely (but linearly) distorting acoustics of the cylinder phonograph system that recorded him. Blind deconvolution is the process of recovering an original signal from a filtered version of it without knowing what filter was used; it works when the filter and the original signal have characteristic natural frequencies (or singularities) in distinct regimes of frequency. The particular algorithm in this case was a form of homomorphic filtering [1], which is a practical example of a real tool that can help implement the regrettably inexact idea of "separating desired from undesired" signals. Stockham, BTW, formed a firm, now DRC-Soundstream I believe, and commercialized the method. People who have heard the deconvolved Caruso asserted to me that the corrupted original sounded better, to their ears (whether because of second-order defects in the process or because Caruso was phonogenic, I won't presume to guess). It may not have been the same case that Paul Dietz was referring to. | I suggest you look at the literature on adaptive filtering. Widrow and | Stearns "Adaptive Signal Processing" in the Prentice-Hall Signal | Processing Series is a good place to start. I agree with this. Unfortunately, classical (Widrow-type LMS-linear) "adaptive" filtering, while it is another very useful tool with a lot of applications, really addresses a different class of problems from what Paul Gallagher proposed. They tend to be applications that can be cast in the form of training a device (the adaptive filter) to give a precisely known output, or to eliminate an interfering signal that is mathematically degenerate in some way (e.g., is narrowband) or is available in some distorted form itself. Again, in every case you need a mathematical, not just an intuitive, basis for attacking the problem. Another basic tool that might have relevance here is dynamic time warping (DTW), used routinely in applications like pattern recognition where inputs (like speech) are subject to uncontrollable time-scale expansions and compressions. Unlike the classes of algorithms mentioned so far, DTW doesn't assume rigid time alignment among the different signals being manipulated and compared. I just mention it. I wish someone with a broad background would attack the problem that Paul Gallagher raised, or at least summarize in the literature the different tools that are steps in that direction. Perhaps it has been done -- if so, please post. The different tools, characteristically, arise in very different fields of inquiry, hence the need for broad background. [1] Oppenheim and Schafer, _Digital Signal Processing_, Prentice-Hall 1975, Chapter 10, is a broad introduction to homomorphic sig. proc. Max Hauser / max@eros.berkeley.edu / ...{!decvax}!ucbvax!eros!max
phd@speech1.cs.cmu.edu (Paul Dietz) (07/16/88)
Max Hauser writes: >In article <2266@pt.cs.cmu.edu>, Paul Dietz wrote: >| In article <4944@husc6.harvard.edu> ... (Paul Gallagher) writes: >| >Why isn't it possible to completely restore a recording: for example, to remove >| >all extraneous noise (hiss, clicks, coughs), even to make a reasonable guess >| >about information not in the original recording (for example, given a score >| >and a knowledge of the harmonics of a voice or an instrument, to recreate >| >something close to the sound of the original performance)? >| Actually, this sort of thing is commonly done. >I'm not sure that I would concur with this. People frequently suggest >strategies similar to Paul Gallagher's, above; but the problem arises in >translating ideas like "knowledge of harmonics of a voice or an instrument" >into hard specifics, algorithms that act on the recorded information, >to do the job. The obstacles are not in the broad concept but in the >nitty gritty. > Unfortunately, classical (Widrow-type LMS-linear) >"adaptive" filtering, while it is another very useful tool with a lot of >applications, really addresses a different class of problems from what >Paul Gallagher proposed. I guess I jumped a head a bit on this on. I was thinking more in terms of devices like noise cancelation systems used in cockpits where LMS and RLS type approaches work quite well. However, I could imagine applying very similar ideas to exactly what Paul Gallagher suggested. For instance, let's say you had a noisey recording of a single trumpet that you wanted to clean up. One approach might be to create a transmission line-like filter (with harmonically related pass areas) that could be adaptively tuned (both in frequency, and peak/valley ratio) to obtain maximal response. Seems to me that this could work pretty well. Of course, trying to do this for more than one instrument at a time instantly becomes a nightmare... (Essentially, this is equivalent to the "narrow band" problem you mentioned LMS was good for...) >Another basic tool that might have relevance here is dynamic time >warping (DTW), used routinely in applications like pattern recognition >where inputs (like speech) are subject to uncontrollable time-scale >expansions and compressions. Unlike the classes of algorithms mentioned >so far, DTW doesn't assume rigid time alignment among the different >signals being manipulated and compared. Sounds interesting! Do you have any references? (Probably one of those books on multirate DSP that I always ignore...) Paul H. Dietz ____ ____ Dept. of Electrical and Computer Engineering / oo \ <_<\\\ Carnegie Mellon University /| \/ |\ \\ \\ -------------------------------------------- | | ( ) | | | ||\\ "If God had meant for penguins to fly, -->--<-- / / |\\\ / he would have given them wings." _________^__^_________/ / / \\\\-
thearlin@vdsvax.steinmetz.ge.com (Thearling) (07/19/88)
>In article (Paul Gallagher) writes: >Here's a question maybe someone out there can answer: >Why isn't it possible to completely restore a recording: for example, >to removeall extraneous noise (hiss, clicks, coughs) > I recall seeing something on TV (PBS?) about the recording restoration lab at the New York Public Library. They have loads of equipment to do just this. From what I remember, they spend most of their time restoring old cylinder and 78 recordings. kurt ----------------------------------------------------------------------- Kurt Thearling thearlin%vdsvax.tcpip@ge-crd.arpa General Electric CRD thearlin@vdsvax.steinmetz.ge.com Bldg. KW, Room C313 uunet!steinmetz!vdsvax!thearlin P.O. Box 8 thearlin%vdsvax@steinmetx.uucp Schenectady, NY 12301 kurt%bach@uxc.cso.uiuc.edu (518) 387-6779 kurt@bach.csg.uiuc.edu -----------------------------------------------------------------------
tody@noao.arizona.edu (Doug Tody X217) (07/21/88)
> >In article (Paul Gallagher) writes: > >Here's a question maybe someone out there can answer: > >Why isn't it possible to completely restore a recording: for example, > >to removeall extraneous noise (hiss, clicks, coughs) It isn't possible to completely restore old recordings given current signal processing technology and algorithms, but the following is a recent development I found interesting. This is a new line of CD's by Phillips called "Legendary Classics". Here is a quote from the liner notes in one of my disks: Historical recordings by such legendary musicians as Ravel, Casals, Prokofiev, Van Beinum, Souzay, Oistrakh, and Richter, now lovingly restored to their original majesty thanks to Sonic Solutions' NoNoise micro-technology and Philips remastering and digital expertise. The Artistry of the Past Now Made Perfect For Philips exclusive Compact Disk Collection of Legendary Classics Well, of course it isn't perfect, but it is the most interesting commercially available development I have seen in this area. The difference in the sound is very noticeable - doesn't mean you will like it though, many people won't. There was a review of the technology in a recent Grammophone... Here is a brief description of the process from the liner notes before me: ...Developed ... at Sonic Solutions, in San Francisco, Calif... NoNOISE can remove surface noise, tape hiss, clicks, and pops and other unwanted noise without affecting the original. ...The digital data is then transferred to large computer disks and engineers at Sonic Solutions assess the problems, using special analytical programmes... If the recording contains clicks and pops, they employ a programme which incorporates techniques developed in artificial intelligence research [I'm just quoting this, don't blame me...]. This identifies the clicks or pops and recreates sound to replace the area they occupied. Since the area is reconstructed rather than removed, the precise duration of the original performance is preserved. To reduce surface noise or tape hiss, ... analyze the spectral composition of the underlying noise floor. Over 2000 points in the audible spectrum are measured to yield an accurate estimate of the unwanted noise. This is then used by a special computer programme to perform "micro surgery" on the sound to reduce the noise without affecting any subtlety of the original. Over 53 million separate computations... [etc.] This is rather overstated, but an interesting experiment nonetheless. I have only one such disk thus far - it is very good sound for such an ancient recording - 1928, 1932, 1938, all mono (the Ravel and Prokofiev disk). Unless you are interested in these ancient performances, however, any well recorded modern performance is going to provide much better sound. Doug Tody NOAO, Tucson
strong@tc.fluke.COM (Norm Strong) (07/21/88)
In article <4867@vdsvax.steinmetz.ge.com> vdsvax!thearlin@steinmetz.UUCP (Thearling) writes: } }>In article (Paul Gallagher) writes: }>Here's a question maybe someone out there can answer: }>Why isn't it possible to completely restore a recording: for example, }>to removeall extraneous noise (hiss, clicks, coughs) }> } }I recall seeing something on TV (PBS?) about the recording restoration }lab at the New York Public Library. They have loads of equipment to }do just this. From what I remember, they spend most of their time }restoring old cylinder and 78 recordings. } A company named Sonic Solutions has come up with a rather complicated computer program for removing noise, etc from old recordings. It doesn't operate in real time; in fact, it takes all night to clean up one side of a 78 at a rate of 54 million computations per second. (sounds fast to me.) What do they call this process? No Noise, of course. -- Norm (strong@tc.fluke.com)
mike@rbdc.UUCP (Mike steel) (07/28/88)
I am looking for a reasonable priced Cd player that will be reliable. I am looking at the Magnavox players. Has anyone heard good or bad about these players? I am generally looking for something in the $200 range. Any sugestions here or via Email would be appreciated. Thanks Mike NETORPRMS at NCSUVM.BITNET