Forums

MELP quality

Started by Brian April 3, 2006
I have listened to MELP vocoder and found that its consonants are 
sometimes too weak, sometimes it produces hoarse speech and clicks, and 
its pitch shakes/vibrates which makes it sounds unnatural.

Is that typical to vocoders at that bitrate (2400 bps)?
Can the above be improved or eliminated?
Can anyone recommend a better vocoder at that rate?

Thanks,
BHD
Brian wrote:
> I have listened to MELP vocoder and found that its consonants are > sometimes too weak, sometimes it produces hoarse speech and clicks, and > its pitch shakes/vibrates which makes it sounds unnatural. > > Is that typical to vocoders at that bitrate (2400 bps)? > Can the above be improved or eliminated? > Can anyone recommend a better vocoder at that rate?
2400bps and hi-fi don't exactly go hand in hand. If it sounds even remotely natural, it isn't too bad for 2400bps. Codecs which sound good enough for wireless phone use, for example, start around 4kbps. iDEN, for example, runs at 4.2kbps. Regards, Steve
Brian wrote:
> I have listened to MELP vocoder and found that its consonants are > sometimes too weak, sometimes it produces hoarse speech and clicks, and > its pitch shakes/vibrates which makes it sounds unnatural.
There are indeed some quality issues like you were describing with the old MELP vocoder.
> Is that typical to vocoders at that bitrate (2400 bps)?
In some vocoders in that range you may encounter such artifacts. But that is not typical to all of them. Some achieve much better quality than the old MELP. Check out the audio demo of Compandent's ComPacketer at 2.8 kbps : http://www.Compandent.com/products_compacketer.htm
> Can the above be improved or eliminated?
Yes, you may use the Enhance-MELP a.k.a. MELPe which achieves better quality than the old MELP for the same bit-rate. Some further quality issues were resolved by Compandent in their MELPe, as you may find at: http://www.Compandent.com/products_melpe.htm http://www.Compandent.com/melpe_faq.htm http://www.melpe.com
> Can anyone recommend a better vocoder at that rate?
Yes, MELPe or ComPacketer by Compandent. Demo of the ComPacketer, Compandent's high quality 2.8 kbps voice coder: http://www.Compandent.com/products_compacketer.htm and demo of Compandent's MELPe at: http://www.Compandent.com/products_melpe.htm
Brian wrote:
> its pitch shakes/vibrates which makes it sounds unnatural. >
You bet it does... All parametric vocoders like MELP require precise estimation of basic parameters of a speech signal - pitch and voicing being the most important. Any errors in pitch and voicing estimation are detrimental to speech qiality. Pitch estimation procedure used by MELP is based on correlation - a rather crude and problematic way to handle this problem. The main problem with all those standards is not that the new and more accurate techniques aren't available - it is that all those big companies want to use their and only their IP in all the standards... Look at what happened to a decade-long effort to standardize a 4kbs toll-quality ITU speech coding standard: it failed miserably because those companies want to have their cake and eat it too... But I am sure that toll-quality speech at 4 kbs-per-second is very much possible, they just need to open standard participation to smaller and more innovative companies. BTW, you shouldn't pay much attention to Compandent's marketing brochures...
>Can anyone recommend a better vocoder at that rate?
To answer your question, try some Spirit products, for example: http://www.spiritdsp.com/alcwi.html Be sure to ask them to sign a patent indemnification clause before you buy anything from them... Let us know if they do...
You should not pay attention to an apparently Russian guy who
unsurprisingly recommended a Russian company... and apparently whose 
only concern seems to be  avoiding IPR payment for hard working 
companies that contributed significant inventions to the field...

On the other hand, Compandent seems to be a very professional company,
that unlike the Russian one has contributed to standards... and does
respect IPR, and apparently developed probably the highest quality
low-rate coders, such as ComPacketer (at 2800 bps):
http://www.Compandent.com/products_compacketer.htm

and provided major improvements to the excellent MELPe vocoder that 
already won several standardization competitions (US DoD, NATO) and was 
professionally tested in numerous conditions, as seen from:
http://www.Compandent.com/products_melpe.htm
http://www.Compandent.com/melpe_faq.htm
http://www.melpe.com

> Be sure to ask them to sign a patent indemnification clause before you > buy anything from them... > > Let us know if they do...
Would you buy a patent indemnification from a Russian company? I'll see you on the dark side of the moon... :) Cheers, USP
>On the other hand, Compandent seems to be a very professional company, > that ... does respect IPR
Good for Compandent... But this also means that their vocoders are not of the highest possible quality... Like I said, can't have your cake and eat it too... As far as some companies selling "proprietory" vocoders sounding better tnan the standard (riding chin-deep in patents) ones... well, it's just funny until it gets sad...
fizteh89 wrote:
> Brian wrote: > >>its pitch shakes/vibrates which makes it sounds unnatural. >> > > > You bet it does... > > All parametric vocoders like MELP require precise estimation of basic > parameters of a speech signal - pitch and voicing being the most > important. > > Any errors in pitch and voicing estimation are detrimental to speech > qiality. > Pitch estimation procedure used by MELP is based on correlation - a > rather crude and problematic way to handle this problem.
Most voice coders do that. However, most developed in the last 15 years then apply another stage of fractional tuning of the result.
> The main problem with all those standards is not that the new and more > accurate techniques aren't available - it is that all those big > companies want to use their and only their IP in all the standards... > > Look at what happened to a decade-long effort to standardize a 4kbs > toll-quality ITU speech coding standard: it failed miserably because > those companies want to have their cake and eat it too...
4.2kbps has been used for toll use in iDEN for years. The codec (unless they changed it) is like the half rate GSM one, with the frames stretched from 20ms to 30ms. It can give pretty reasonable quality when done right.
> But I am sure that toll-quality speech at 4 kbs-per-second is very > much possible, they just need to open standard participation to > smaller and more innovative companies. > > BTW, you shouldn't pay much attention to Compandent's marketing > brochures...
Yeah. The Compandent marketing department gets a bit tiresome every time someone mentions MELP here. Regards, Steve
fizteh89 wrote:
> But this also means that their vocoders are not of the highest possible > quality...
Right, and in order to get that reality you needed to simply stop whining and go to their web site and listen to the impressive CompPacketer's audio demo: http://www.Compandent.com/products_compacketer.htm
>From Compandent's web-site (http://www.compandent.com/melpe_faq.htm):
"10.1 Q: Suppose I developed MELPe implementation based on the standard, or obtained such from some source other than Compandent and I use it commercially, do I still need to pay any royalty to Compandent? A: Regardless of where you got your MELPe implemenation from, the answer is yes!, if your implementation uses our IPR in the MELPe standard reference. Compandent has made numerous contributions to the MELPe standard reference. We reserve all our rights in it. We can provide you a complete list of our contributions. Any party intending to develop products based on MELPe should contact Compandent (as well as other IPR holders) regarding licensing terms and conditions." As we already know (from last discussion here), Compandent doesn't have any patents on the MELPe standard. And copyright protection does not work in case of a clean-room engineering. So, what IPR can they possibly have in MELPe to be entitled to any royalties ? Trade secrets ? Ha-Ha-Ha-Ha-Ha....................