Audio Recognition issues 7/3/23

🎧 Audio Recognition
1

Hey,

I assembled a 2 hour prerecorded show for KRVM-FM that ran from 5-7 PM on July 3rd. Link to the archive and playlist as logged by Spinitron is here:

Select Vinyl Revival at 5PM on July 3rd.

The show went off without issues (except for being 2 mins short, my fault) but audio recognition was very spotty. Between 6:19 and 6:49, it missed 6 songs including 5 in a row. Cream “I Feel Free” and Beau Brummells “Laugh Laugh” for example. Rickie Lee Jones was skipped around 5:23 and Thin Lizzy around 5:55.

Was there a technical issue at the time? Or is this normal. I expected it to be stumped by a couple obscurities by The Doors and ZZ Top but it ID’d them fine.

Thanks for any insight.

Oh yeah, I have no idea where DJ Kaos came from at 5:13.

Dave

2

Given the name of the show I assume you were playing off a turntable. Is that so?

We see this sort of thing occasionally (note the occasionally) and are wondering if it’s a calibration issue with the turntable(s), some variation in the recording(s), or both.

In tests we are able to fly in under the recognition radar by changing the pitch of a track by a few percents, hardly enough to notice by ear but enough to alter the fingerprint to avoid detection.

You can test recordings and tracks here. I believe the site is operated by the same folks (ACRCloud) that do the audio recognition for Spinitron.

If you do any tests I’d love to know what your results are.

3

This show was recorded from a single Technics SL-1200MK7 turntable. I wouldn’t do the show for the station at all if I didn’t have audio gear that produces excellent audio quality. The pitch is dead-on - I can’t stand it when a turntable’s speed isn’t perfect! :wink:

The recordings are CD-quality WAV files recorded with Audacity. Strict attention is paid to making the best quality recordings possible without going beyond the needs of FM radio. I’m more than a little OCD about this.

Each hour of the show consists of four segments (WAV files) of approximately 20, 15, 15, and 10 minutes. The eight segments for a 2-hour show are submitted to the station as WAV files and the station drops them in their automation system (Simian) in WAV format. So they’re basically untouched from recording to airing with no re-encoding or conversion to a lossy format.

The regular host of the show does it live in the studio but he couldn’t make this date, so I did the recording and assembly for him (he provided voice tracking). Normally he enters his playlist into Spinitron by hand. We decided to try song recognition this time and were kind of disappointed. That’s why I’m wondering if there was a technical issue - note the insertion of a song we didn’t play, during another song that was correctly identified.

Thanks for your suggestion but it definitely wasn’t an issue with pitch, or the quality of the recording.

-dave

4

I played back the entire Spinitron archive of the show and it’s fine. No dropouts or other anomalies. I guess this will remain a mystery.

5

I want to make sure I understand you correctly. You’re saying you replayed the entire show in the normal fashion and it went off without a hitch? There was no problem with recognition?

6

Hi @xdrguy iiuc you record the show in advance. I guess that recording is put into automation software which plays it out at a given time.

Check 1) Some show producers and/or stations adjust the duration of a finished program to make it fit a time slot by applying a small time stretch or pitch shift. With modern digital tools you’ll be hard pressed to notice a 60 sec adjustment to a 58 min audio file. We’ve experience where small adjustments affected recognition performance a lot.

Check 2) Are the recordings you played fingerprinted in the ACRCloud database? You can check using https://aha-music.com/ which, iiuc, use the same database.

If 1) and 2) don’t explain the omissions we need to dig deeper.

Check 3) You have the finished audio files of the shows you make. Check those using https://aha-music.com/ If recognition is not working on that unmolested audio file then something in your DJ rig or digital recording methods is to probably blame.

(Using the strobe on the SL-1200 is probably good enough to check its speed. The “pitch” slider being centered isn’t as reliable.)

But there’s often a lot of signal processing at the station before the audio you produced gets into the live stream, which in KVMR’s case is mp3, 44100 Hz, mono, fltp, 64 kb/s. That nasty encoding is probably after limiting and may even come off an FM monitor. But this stream is what ACRCloud uses for input to recognition. So we want to check that too in the next step.

Now, KRVM is using the Spinitron archive so if 1), 2) and 3) all check out and everything is recognized in your clean audio files then…

Check 4) Play the archive into https://aha-music.com/ and see what happens.

If everything in the stream archive of your show is recognized in Aha-Music then the next step is to give the URL of the archive stream to ACRCloud with the list of recognitions (timestamp, artist, title) you expect and have them check it. The URL for you will look like this

https://ark3.spinitron.com/ark2/KRVM-20230712T150000Z/index.m3u8

but replace 20230712T150000Z with the date and time of the start of your program in UTC (GMT) time zone. The date time code in that example is for 2023 July 12 at 15:00 UTC, which formatted numerically with seconds is 2023 07 12 at 15 00 00 UTC and we use T to separate date and time and Z to mean UTC and remove the spaces.

7

Although the show was 2 mins short as submitted, the station tacked on a 2 minute song to fill it out. No time stretching or anything.

At least a couple of the songs are not fingerprinted, or at least not recognized when I upload clean 320kbps MP3 files. So we know there will be some gaps. However, well-known songs by Cream and Beau Brummells are recognized using this method, but weren’t ID’d from the stream. Anyway there seems to be a daily limit on how many songs I can submit for analysis. I used up my quota.

Thanks for your helpful suggestions. I’ll continue to play around with this as time allows.

-dave

1 Like
8

Ok. We can always escalate with a request for technical analysis by ACRCloud so long as we provide audio and a timestamped list of what we expect it to recognize.

I didn’t know Aha-Music had a quota. Makes sense.

9

No, I have not played the entire 2-hour archive back to check song recognition. Today I did upload some selected songs from very clean 320kbps MP3 files that were not recognized.

Thin Lizzy “Sweet Marie” Taj Mahal “Farther On Down The Road” Alice Stuart “So Free”

That’s where I had to stop due to using up my daily quota.

I fully expect a few obscure songs from a given playlist to not be ID’d. But Cream “I Feel Free” and Beau Brummells “Laugh Laugh” we’re not ID’d from the stream, but are ID’d when uploaded individually.

Based on this limited testing, the host of the show is going to continue entering his playlists manually. He doesn’t like anything not being ID’d in his Spinitron playlist.

Thanks again everyone for your help.

-Dave

1 Like
10

Agreed, @xdrguy, I’d expect those to be recognized too.

So if we can get a URL to the archive and the timestamp/artist/title of the items what were not recognized then we have something that we run in a lab test.

11

These are all the spins in the archive from the July 3rd show “Vinyl Revival” that were not recognized. Hope this gives you something to work with!

Rickie Lee Jones “Night Train" https://ark3.spinitron.com/ark2/KRVM-20230704T002334Z/index.m3u8

Thin Lizzy “Sweet Marie” https://ark3.spinitron.com/ark2/KRVM-20230704T005540Z/index.m3u8

Bruce Springsteen & E St Band - “4th of July, Asbury Park (Sandy)" LIVE https://ark3.spinitron.com/ark2/KRVM-20230704T005957Z/index.m3u8

Cream “I Feel Free" https://ark3.spinitron.com/ark2/KRVM-20230704T012144Z/index.m3u8

Taj Mahal “Farther On Down The Road” https://ark3.spinitron.com/ark2/KRVM-20230704T012822Z/index.m3u8

Alice Stuart “So Free” https://ark3.spinitron.com/ark2/KRVM-20230704T013416Z/index.m3u8

David Bowie “Lady Grinning Soul” https://ark3.spinitron.com/ark2/KRVM-20230704T013822Z/index.m3u8

Thunderclap Newman “When I Think” https://ark3.spinitron.com/ark2/KRVM-20230704T014156Z/index.m3u8

Beau Brummells “Laugh Laugh” https://ark3.spinitron.com/ark2/KRVM-20230704T014503Z/index.m3u8

1 Like
12

Hi @xdrguy Wow you provided a URL for each song, that’s great and should make it even easier.

@Peng_from_ACRCloud Hi Peng @xdrguy is curious to know what might explain the poor recognition rate on some of his programs on KRVM. Iiuc, each item in the list in the above post is a song that ACRCloud didn’t recognize or one that had a score of below 60, which is the minimum that Spinitron applies. Can you determine what might explain the low hit rate?

Technical note, KRVM’s live stream that ACRCloud uses is mp3, 44100 Hz, mono, fltp, 64 kb/s. The URL’s listed above are HLS streams that we make for an on-demand replay service of the live-stream. They have been trans-coded to AAC for better player compatibility.

1 Like
13

Hi Tom and Dave,

We noticed the problem is caused by the pitch/speed change mostly from the orignial track. We now have enabled to detect them. Only this song called Alice Stuart “So Free” we can not find from our database or elsewhere on the market. https://ark3.spinitron.com/ark2/KRVM-20230704T013416Z/index.m3u8

1 Like
14

Hello @Peng_from_ACRCloud ,

Thanks for writing back so quickly.

Are you saying that you have applied a general fix for tracks with pitch/speed changes or just for the tracks in question?

-Thanks again

15

@Peng_from_ACRCloud @tom

Thank to everyone who has helped look into this.

Is the pitch problem related to transcoding KRVM’s mono MP3 stream for the archive? Is it inherent in their stream? Or do you think the problem is with my WAV source files? I can provide links to the master WAV recordings for these songs if that would help.

Would it help if you were monitoring KRVM’s 96kbps stereo AAC+ stream, rather than their mono 64kbps MP3 stream?

https://c1865s644.fastserv.com (preferred)

http://lin2.san.fast-serv.com:6002/stream (This link is probably deprecated and shouldn’t be active)

Note that I’m a KRVM volunteer, not an official spokesperson for the station. :slight_smile:

Thanks again for all the help with this.

-dave stambaugh

16

Audio codecs don’t affect pitch, afaik. I mentioned above the places in the production and playout chains that are in my opinion mostly likely culprits. Narrowing in on the actual culprit is a local engineering task. There’s only so much I can contribute with offering guesswork.

17

Thanks Tom & Peng, appreciate your time and input on this. I might do a little more detective work at my end.

-dave