Removing adds from podcasts

Removing adds from podcasts

Aug 7, 2020

I am sick of the psyops of advertising. I want to detox. but I still want to listen to podcasts. I need strip out the adds from them

How can I programatically remove adds from podcasts

  • most podcasts have an add role at the beginning, end or middle
  • the Beginning and end are easy enough to skip, it’s built into my podcast player, to skip X seconds. I use this to avoid annoying jingles and “brought to you by” messages.
  • middle bits might have silence or a tune to signal their coming.
  • would want a way to identify the timestamps of the beginning and end of this, and then, skip it real time (hard without modifying a player app)
    • or clip it out and re-save the podcast.
      • if I save it it raises the question. Where to, and for how long
  • I am thinking I could make this a lambda function, feed by the rss of my podcasts, that resaves to an object storage provider, and makes a new rss feed, that I can subscrice to with my app.
  • First I have to research about how to detect middle bits.
    • it could be that they are always at the same time in the more reliable podcats.
  • sometime there is info in the middle bit I want to hear too, so maybe .. ASR it to text, and add the text to the feed description. It is not so invasive to see the text.. then I can also regex out square space and blue apron and things.

Notes

dpwe/audfprint

Dpwe’s audfprint Is a python command line tool for working with audio landmarks. It seems to be a reimplementation of a matlab project.

Install

  • pip(3?) install -r requirements.txt
    • make sure you have ffmpeg.. the OSX ffmpeg might give slightly different output (i.e time formatting) to the linux one. It doesn’t seem to be a problem.

Use.

  • Get some podcasts. (in this use case the podcast play a little tune(motif) at the beginning and end of the ads. beginning tune and end tune differ)
  • Clip out the audio motifs that bookend the middle bit with adds. This can be done with audacity’s export selection. Wav files seem to be fine (anything ffmpeg can open?).
    • save these clips to named for the start and ends.
  • Make a landscape database for the start motif and the end motif.
    • python audfprint.py new –dbase podcast_middlebit_in.pklz podcast_clip_in_*.wav
    • python audfprint.py new –dbase podcast_middlebit_out.pklz podcast_clip_out_*.wav
    • Use this DB to scan podcast to identify the position of the start and end of the middle bit.
      • OSX will output in -1500.9s negative seconds format
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

bash-3.2$ python3 audfprint.py match --dbase taz_fpdbase.pklz ../audio/taz14_tc.mp3 --find-time-range -T -x 10
 -H 2
Mon May 18 16:42:36 2020 Reading hash table taz_fpdbase.pklz
Read fprints for 4 files ( 1262 hashes) from taz_fpdbase.pklz (0.00% dropped)
Mon May 18 16:43:49 2020 Analyzed  ../audio/taz14_tc.mp3 of 4756.956 s to 298377 h
ashes
Matched    1.3 s starting at 1932.3 s in ../audio/taz14_tc.mp3 to time   18.5 s in /Users/bwp002/Documents/taz
_13_mid_in_audfprint.wav with    10 of   216 common hashes at rank  2
Matched    2.7 s starting at 1931.8 s in ../audio/taz14_tc.mp3 to time    7.9 s in /Users/bwp002/Documents/taz
_mid_in_audfprint.wav with    25 of   314 common hashes at rank  0
Matched    1.5 s starting at 1932.8 s in ../audio/taz14_tc.mp3 to time    7.2 s in /Users/bwp002/Documents/taz
_13_mid_in_audfprint.wav with    16 of   216 common hashes at rank  2
Matched    1.7 s starting at 1932.3 s in ../audio/taz14_tc.mp3 to time    3.8 s in /Users/bwp002/Documents/taz
_mid_in_audfprint.wav with     8 of   314 common hashes at rank  0
Matched    0.8 s starting at 1934.3 s in ../audio/taz14_tc.mp3 to time    5.4 s in /Users/bwp002/Documents/taz
_mid_out_audfprint.wav with    13 of   197 common hashes at rank  1
Matched    4.0 s starting at 1932.3 s in ../audio/taz14_tc.mp3 to time    2.5 s in /Users/bwp002/Documents/taz
_mid_in_audfprint.wav with   109 of   314 common hashes at rank  0
Matched    2.5 s starting at 1932.8 s in ../audio/taz14_tc.mp3 to time    2.5 s in /Users/bwp002/Documents/taz
_mid_out_audfprint.wav with    61 of   197 common hashes at rank  1
Matched    1.5 s starting at 2075.1 s in ../audio/taz14_tc.mp3 to time   17.6 s in /Users/bwp002/Documents/taz
_13_mid_in_audfprint.wav with     7 of   216 common hashes at rank  2
Matched    2.9 s starting at 2075.1 s in ../audio/taz14_tc.mp3 to time    7.5 s in /Users/bwp002/Documents/taz
_mid_in_audfprint.wav with    24 of   314 common hashes at rank  0
Matched    1.2 s starting at 2076.5 s in ../audio/taz14_tc.mp3 to time    7.2 s in /Users/bwp002/Documents/t
az_13_mid_in_audfprint.wav with    10 of   216 common hashes at rank  2

TODO Use start and stop timestamps to clip out and re-compile podcast with no-middle bit.

Ref

1
2

bash-3.2$ ffmpeg -i audio/taz14_tc.mp3 -filter_complex "atrim=start=0:end=1932[s];[0:a]atrim=start=2075[e];[s]
[e]concat=v=0:a=1[af]" -map [af]  out.mp3

which outputs this:

Output mp3

But this is slow as it has to re-encode the file. If I try the “-acodec copy” flag, it says it doesnt work with complex filters.

TODO Make this .. an automated system

  • Make a DB of IN/OUT motifs for podcasts that you listen to