Envelope of Speech

I need to get a smooth as possible envelope of a real-time speech
signal.
I have tried rectification + filtering and squaring + filtering and
also a Hilbert Transformer (to creat the imaginary part of a complex
waveform which I can take the magnitude of)  but neither of these
methods appear to give a smooth envelope. Is there a tried and tested
best method?


Hardy

HardySpicer wrote:

> I need to get a smooth as possible envelope of a real-time speech
> signal.

Main question: what are you trying to accomplish by that?

> I have tried rectification + filtering and squaring + filtering and
> also a Hilbert Transformer (to creat the imaginary part of a complex
> waveform which I can take the magnitude of)  but neither of these
> methods appear to give a smooth envelope.

What is the definition of "smooth envelope" ?

> Is there a tried and tested
> best method?

It depends on how much of the processing delay wrt realtime is 
acceptable. You have to overlook at least one period of the lowest 
pitch; that is about 20ms.


Vladimir Vassilevsky
DSP and Mixed Signal Design Consultant
http://www.abvolt.com

HardySpicer wrote:
> I need to get a smooth as possible envelope of a real-time speech
> signal.
> I have tried rectification + filtering and squaring + filtering and
> also a Hilbert Transformer (to creat the imaginary part of a complex
> waveform which I can take the magnitude of)  but neither of these
> methods appear to give a smooth envelope. Is there a tried and tested
> best method?
> 
> 
> Hardy
> 

Had a similar problem, but using stored data. If you can tolerate a 
delay comparable to the lowest frequency present.

Detect time of sign change derivative of absolute value of signal.
Use that "time" to sample the input signal, interpolate/filter between 
those points as desired.

My experiments were related to following amplitude of an AM modulated 
carrier. I was thing along lines of a peak follower. I may have my 
Scilab code with a demo test case.

On Sat, 4 Oct 2008 13:00:50 -0700 (PDT), HardySpicer
<gyansorova@gmail.com> wrote:

>I need to get a smooth as possible envelope of a real-time speech
>signal.
>I have tried rectification + filtering and squaring + filtering and
>also a Hilbert Transformer (to creat the imaginary part of a complex
>waveform which I can take the magnitude of)  but neither of these
>methods appear to give a smooth envelope. Is there a tried and tested
>best method?
>
>
>Hardy

Hi Hardy,
   My guess is that if you filter out all 
of the signal's noise, that's outside the 
signal's bandwidth, and perform the 
Hilbert process to obtain a time-domain 
"magnitude" sequence, that this should be 
about your best option.

I'm not exactly sure what you mean by "smooth", 
but is it possible that your computations are 
correct and that your signal's envelope is just 
not "smooth"?

[-Rick-]
  

HardySpicer wrote:
> 
> I need to get a smooth as possible envelope of a real-time speech
> signal.
> I have tried rectification + filtering and squaring + filtering and
> also a Hilbert Transformer (to creat the imaginary part of a complex
> waveform which I can take the magnitude of)  but neither of these
> methods appear to give a smooth envelope. Is there a tried and tested
> best method?

Best method for doing what exactly? For instance, you might be trying to create
a display so that 5-10 seconds of speech can be viewed at once.

-jim


----== Posted via Pronews.Com - Unlimited-Unrestricted-Secure Usenet News==----
http://www.pronews.com The #1 Newsgroup Service in the World! >100,000 Newsgroups
---= - Total Privacy via Encryption =---

On Oct 5, 1:00&#4294967295;am, HardySpicer <gyansor...@gmail.com> wrote:
> I need to get a smooth as possible envelope of a real-time speech
> signal.
> I have tried rectification + filtering and squaring + filtering and
> also a Hilbert Transformer (to creat the imaginary part of a complex
> waveform which I can take the magnitude of) &#4294967295;but neither of these
> methods appear to give a smooth envelope. Is there a tried and tested
> best method?
>
> Hardy

take chunks of say 20ms of speech convert to frequency domain and
apply LPC
analysis in frequency domain as opposed to time domain.You will get an
envelope
of speech in time domain. But this works well only for voiced segments
of speech.