Minimum sample size of PCM stream

[Justin Haygood]

I am evaluating using EchoPrint for a project using a different set of data samples than a music database (private store of audio data to match against). Does the code generator need a full 20 seconds to create a fingerprint, or can a smaller (1-5 second) sample work as well?

[Andrew Nesbit]

Hi Justin,

The codegen will create accept any length of audio, but you need at least several seconds before it will generate any hash codes at all. For reliable matching we recommend at least 20 seconds.

Best,

Andrew

[Justin Haygood]

Thanks. Many of our audio samples will be shorter than 15-30 seconds and there won't be a huge chunk of samples.. Ill be testing the matching with 2 and 3 second samples this week to see how well it performs.

[Mario Rossi]

Hi Justin, i have same project,  have you some suggestion to share for 3-5 sec jinle song recognition?

thank you very much cristiano...@hotmail.com

[Mario Rossi]

Hi, i'm looking an echoprint expert for big international project.

I need to build on amazon a media monitoring solution with echoprint that will be able to accept any broadcast source like tv streaming , radio streaming etc.. and match with my DB of fingerprint.

thank you

Cris

On Monday, June 11, 2012 4:48:18 AM UTC+2, Justin Haygood wrote:

[Justin Haygood]

Hello Mario!

3-5 seconds might be to short, but there's a few changes we did to make it better for short (15-30) content with OTA recognition. I've attached 2 patches that we used that made it work a lot better, however, fingerprints generated aren't compatible with stock libcodegen (however, reference fingerprints created using the patched libcodegen matches sample fingerprints created using the patched libcodegen).

The changes basically involve:

1. Increasing density of fingerprints (more data in less time) to increase match rate for short samples
2. Reducing false positive rate by skipping over "empty" hashes.

However, we've seen much better match rates (though 3-5 seconds is probably on the short side even then) with some of the proprietary ACR technologies. I highly recommend you take a look at them.

--
You received this message because you are subscribed to a topic in the Google Groups "echoprint" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/echoprint/J7MQftCfpCM/unsubscribe?hl=en.
To unsubscribe from this group and all its topics, send an email to echoprint+...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.
 
 

[Mario Rossi]

Thank you very much Justin, regarding proprietary ACR technology i konw very well one of them...

 i'm seriously evaluating to leave "proprietary solution" and to build my solution with echoprint technlogy because licence make no sense in big implementation(too much cost).

Best,

Mario

[Mario Rossi]

Andrew i need urgently to contact you, is possible to have your email ?

[cheche]

Hi,man I now try to modify code_gen but it is too hard ,do you have some Documents/wihte-paper?

Thank you.

在 2013年4月6日星期六UTC+8上午7时42分16秒，Justin Haygood写道：

[Andrew Nesbit]

On Thu, Jun 27, 2013 at 1:06 PM, cheche <chehon...@163.com> wrote:

Hi,man I now try to modify code_gen but it is too hard ,do you have some Documents/wihte-paper?

Check the link in the other reply I sent you.

Andrew

[cheche]

Hi Andrew:

    Mr Justin Haygood had give us two patches. I try it ,the score had improve .So why didn't you to add them to your codes? May be it will have some harm I don't know?

Thank you .

[车宏霖]

Hi Andrew:

    Mr Justin Haygood had give us two patches. I try it ,the score had improve .So why didn't you to add them to your codes? May be it will have some harm I don't know?

Thank you .

Minimum sample size of PCM stream

Justin Haygood

Apr 6

Hello Mario!

3-5 seconds might be to short, but there's a few changes we did to make it better for short (15-30) content with OTA recognition. I've attached 2 patches that we used that made it work a lot better, however, fingerprints generated aren't compatible with stock libcodegen (however, reference fingerprints created using the patched libcodegen matches sample fingerprints created using the patched libcodegen).

The changes basically involve:

1. Increasing density of fingerprints (more data in less time) to increase match rate for short samples
2. Reducing false positive rate by skipping over "empty" hashes.

However, we've seen much better match rates (though 3-5 seconds is probably on the short side even then) with some of the proprietary ACR technologies. I highly recommend you take a look at them.

- show quoted text -

- show quoted text -

-- 
You received this message because you are subscribed to a topic in the Google Groups "echoprint" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/echoprint/J7MQftCfpCM/unsubscribe?hl=en.
To unsubscribe from this group and all its topics, send an email to echoprint+...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.
 
 

Attachments (2)

changeset_SecondScreen_trunk_Libraries_libcodegen_r51612.diff

774 B   View   Download

changeset_SecondScreen_trunk_Libraries_libcodegen_r51391.diff

836 B   View   Download

--

You received this message because you are subscribed to a topic in the Google Groups "echoprint" group.

To unsubscribe from this topic, visit https://groups.google.com/d/topic/echoprint/J7MQftCfpCM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to echoprint+unsub...@googlegroups.com.

[Andrew Nesbit]

On Tue, Jul 9, 2013 at 12:12 PM, 车宏霖 <chehon...@163.com> wrote:

    Mr Justin Haygood had give us two patches. I try it ,the score had improve .So why didn't you to add them to your codes? May be it will have some harm I don't know?

The ideas are good but I have two reasons for not including these patches "as is":

1. They will introduce incompatibilities with the current stable version (4.x) of the codegen. We should give some thought as to how multiple incompatible versions of the fingerprints can co-exist in the Echoprint ecosystem, without causing confusion.

2. Increasing the density of hash codes affects scalability in the number of songs that can be searched in a reasonable time on the server. We are working on methods for effectively increasing the hash code density whilst maintaining scalability of the searching performance.

Andrew

[Andrew Nesbit]

Note that if you are running your own Echoprint server, then the above points are not necessarily such a big problem.

Andrew

[cheche]

HI Andrew：

    As you say I test it in my own server.Up to now it's ok .

    Now another question in your code segmentlength = 60 * 1000.0 / 23.2 . When I use new codegen ,the code-length will  grow and the segmentlength should change? If change ,hwo to set it ?

thank you.

在 2013年7月9日星期二UTC+8下午7时46分31秒，Andrew Nesbit写道：

[cheche]

Hi 

I used your patches.Thank you very mutch .

Now I have a question ,in old code segmentlength = 60 * 1000.0 / 23.2 . When I use new codegen ,the code-length will  grow and the segmentlength should change? If change ,hwo to set it ?

thank you.

在 2013年4月6日星期六UTC+8上午7时42分16秒，Justin Haygood写道：

[Andrew Nesbit]

On Thu, Jul 11, 2013 at 5:46 AM, cheche <chehon...@163.com> wrote:

Now I have a question ,in old code segmentlength = 60 * 1000.0 / 23.2 . When I use new codegen ,the code-length will  grow and the segmentlength should change?

That means the segment length is 60 seconds. You don't necessarily need to change it.

Andrew