Freesound Forums

Dear Freesounders,

as part of our research for improving Freesound we are developing a tag recommendation system to help Freesound users tagging the sounds when they are uploaded.

Before implementing this system into Freesound we have started a testing phase in which we will compare several strategies for tag recommendation that we have developed and finally choose the one that works best according to your input.

For that, we kindly ask you to participate in an experiment we have set up at http://labs.freesound.org/tagrecommendation/. Please, read carefully the instructions and follow all steps until the experiment is finished. It will take around 20 minutes.

Once we close the experiment, we will randomly select two of the participants and send a Freesound t-shirt as a reward!

Thank you very much in advance for your participation. We believe you'll enjoy the experiment and will be of invaluable help for improving Freesound!

You can reply to this forum post if you have suggestions or ideas...

It was nice to do the experiment. Hope it helps. Will do it again. :-0

Hi Frederic.
Interesting experiment. It makes one aware of how difficult it is to describe/tag appropriately.
I'm afraid there were many sounds (especially of music or synthesized sequences) that I'm quite ignorant about and would not upload or search for - I was at a loss how to tag effectively. I don't know the differences between dub, dub-step, techno, garage, trance, electro, etc. etc., or the type/names of many instruments used etc. so I hope my input on some of these will not affect the results adversely.
Since different users will be interested/knowledgeable in quite different aspects/types of sounds (I'm more into foley and natural recordings) it might be helpful if participants could skip sounds that leave them clueless. Perhaps they can, in which case it might be helpful to point this out in the instructions. I did not think of this until I'd completed the experiment.

In FS1 tags were arranged alphabetically, which was very helpful if scanning through a long list of tags. In FS2 they seem to be more random. I think it would be most helpful if someone tagging their sounds (or indeed filling in a search box) could put tags in some order to prioretise relevance.

Another area that complicates the tag searching process is too many not-quite-relevant tags. For instance, as I remember somebody pointing out, looking for an "explosion" gives 1064 results, most which seem to be "farts" and not really explosions at all. Even though one can "-fart" to remove them from the results we cant "-this" "-that" and "-the-other" to remove everything that's irrelevant so I think we should be encouraged NOT to put in too many "catch-all" tags if an obvious tag will do all the catching. I think a fart recording barely needs more than one tag, i.e. "fart". In retrospect I've been guilty of over zealous tagging myself, putting a "blackbird" tag in a recording of wind in the trees. The actual bird is a twenty millisecond tweet in an 8 minute recording! I later deleted the tag and just left the bird reference in the description.

Anyway, just my thoughts to provoke further discussion.

Wibby.

tip!!

if I had to write an algorithm for tagging-help I should ask the user to type 'the three most important' tags themselves and started to suggest new tags from that point. Maybe it pointed the system in a serious way more easy.

hope you understand what I mean.

Boy, you have no idea of how difficult is this test for us non-native speakers... Even for field-recordings - in which I have some expertise, I always have to resort to en English dictionary, but these synthetic sounds that simply don't suggest ANYTHING to me... Makes me feel like an idiot, just hate it! I went through the test once, but don't expect me to go through it again, no way
D

Ha ha Haa.. dobroide, you really made me laugh!

I AM a native speaker yet I have total sympathy for your position. My shortcoming is that I have no knowledge of the "language of music", especially the modern stuff. (I use the excuse of old age and an unreliable memory.) Thankfully I have faith the guys who are sorting this tag thing will forgive us both.

Wibby.

dobroide,
As a non-native speaker I have tree options for tags.
1] My wife
2] A list of tags I have made in exel the last few years.
3] For eg a field-recording I go to freesound and search on 'field-recording' and one ore two hot items in the recording. Than I play with the minimum time (<120sec.)

The result of the others I use for searching for the best tags.

Hope it helps you to.

When you send me a PM with email and your language I can send you my field-recording tags.

Ah ah ah - everyone seems to have problems with synth sounds. And not only that, some of the sounds on my test were recordings (but of what?).

Just occurred to me...
Synths simply show this in a more obvious way. But there are tags that only the recorder or creator of the sound can possibly add - What gear/equipment/software was used? Place. Time of day. Date / time of year.

A recording of a clanging, banging, snapping noise can be just as puzzling. Was it made by a kitchen appliance, a door, a machine or something else ?

Other things trip you over as well. Unless you are a trained musician and/or are using a metronome when you listen to the sounds, can you tell tempo signature and BPM? These are often used as tags and very useful for someone trying to incorporate those into their own music.

It's great that this is in the works. Thanks guys for working hard to make freesound better.

At this point most of the suggestions by the system were weird, almost random. I guess it's hard to make it work well.

So im tagging the sound and in the tag recommendation it brings up "urination"....... hahaha tha hell?

Its really difficult tagging sounds were the context is unknown.

It works relatively good with large reference database, and correctly created one (depends on user's experience). Generally, with larger variety of subjects - it reflects "collective consciousness/thinking" of humankind. Don't ask, why I thought about Homer Simpson.

Maybe google/youtube could help? I.e. maybe they have some database to share? As far I remember - they like to interact with other projects from time to time, because it's good for their image.

That was exhausting Is it just me or do others also need most often more time to do surveys as were estimated?

I find it hard to tag my own sounds, but tagging others is much more difficult. But in huge databases it's crucial to find what you are searching for. So I appreciate your efforts to improve sound tagging.

I would like to know whether the suggestions will be based on the words I typed till then only, or are there efforts to algorithmic analyze the sounds. I.e. today, it's not magic anymore to analyze the bpm of music, especially drum loops or the like.
Or maybe one could analyze the sounds for patterns (data mining; spectrum, pitch, time, ...) and suggest common tags out of the set of sounds that match the same or similar patterns.

Just some thoughts ...

Putting a reply in here because I want to be kept up to date on Frederic's experiment and feedback!

- bram

I'm not having that much problems with it. I can find tags to put in easily. If you're struggling, maybe try identifying the sound first. Is it a field recording, foley, synth or 'experimental'? That already gives you one tag. Then just describing what you hear.

I have to add, I did the experiment a few times now, and I've had a couple of samples twice. Don't know if that's very lucky or if that means the sample pool is just quite small.

I would like to participate in the experiment, but no music plays for me to tag. I've tried everything, and nothing works. I would appriciate if some one could give me some suggestions.

Oh! One quick thing: If you're reading this, I have two other forums open, both with no reply. I stated a few questions and I really need them anwered. So, if you could, please check out (and hopefully answer) my realativly short questions onm those Forums. I'm saying this on every post I make so I can raise awairness, but nobody seems to notice. So, please, answer my questions. The name of the Forums are as Follows: "Some One! Please Answer MY Questions!!!" and "(Hidden field recaptcha_challenge_field) This field is required?" They are both listed under the "Bug Reports, Errors, and Feature Requests" section.

THANKS A MILLION,
Spoonandlessspoons.

I would like to say users who upload sounds to not upload anything that has long amounts of silence.

It uses up server space and no one wants to sit there and wait for the sound to play, makes you think that something is wrong with the site, the sound file or your connection.

Also it would be a big plus to have some sort of sound normalization system so that any sound that you click on will play at the same level.

This keeps from having to constantly raise and lower the volume on your system every time you select a sound because this one is very low and the next one blasts the hairs off your head.

Please no normalization, it won't solve anything. You don't want calm well recorded atmosphere to be blasted to 0dBfs. Anyway if something is really loud, you can see it on the waveform and turn your volume down. There is also no point of making quality samples louder just to conform to some distorted/clipped crap.

matucha wrote:
Anyway if something is really loud, you can see it on the waveform and turn your volume down.

A simple, elegant and intelligent advice.
A quick glance at the waveform will tell you if it is safe to press play or if your speakers are likely to spit out their magnets.

Hello everyone,

this post is to announce that the experiment is finished now so no more participants are allowed.
Thank you very much for your participation!

I'm currently analysing the results so I can get an idea of what works and what does not work from the tag recommendation systems being tested. Just for your interest, I'll explain that there were 4 different tag recommendation systems. Two of them were random systems which did not take into account the tags already entered and simply recommended a list of randomly selected tags from a big pool. I was expecting that these recommendations would not be useful at all. The other two tag recommendation systems were more advanced and based on finding similar tags to the ones already introduced. So far, the results I've analysed go along with the expectations and one of the four methods seems to be clearly more useful than the others.

All the comments you entered are very useful also to understand the difficulties of the tagging process. As many of you point out, it is really hard to describe some sounds if you did not record them. I was aware of that problem and I'm aware that the tag recommendation system only helps in some specific situations (ex: current tag recommendation won't help in tagging the pitch of an instrument note recording). Nevertheless, there are some typical tagging problems which could be addressed in different ways and this experiment is being really helpful to understand them.

Again, thank you all for your participation!
For those interested, I'll post here a deeper report of the results once I have it ready (this will be in a few months).

edit:
And by the way! I've already contacted the winners of the t-shirts and sent them