[FFmpeg-user] Advice on using silence removal

[Alex R]

Hi everyone,

I am attempting to leverage ffmpeg in a project that involves recording
short audio clips. So far I have gotten some mixed results and I'd like to
tap into your collective knowledge to ensure my approach is sound.

Context:
- a person records an audio clip of themselves pronouncing a word (imagine
that you read aloud a flash-card that says "tree" or "helicopter")
- the recording is usually made on a mobile phone

The clip contains some silence at both ends, because there is a delay
between the moment the user presses the record button, the moment they
pronounce their word, and the moment they press "stop". Depending on the
device, there may also be an audible click in the beginning.

My objective is to trim the silence at both ends and apply fade-in/out to
soften the clicks, if any.

The challenges are:
- ffmpeg's silenceremove filter needs a threshold value, however,
- each user is in their own environment, with different levels of ambient
noise
- each device is unique in terms of sensitivity

Thus, I can achieve my desired result with one specific clip through trial
and error, tinkering with thresholds until I get what I need. But I cannot
figure out how to detect these thresholds automatically, such that I can
replicate the result with a broad range of users, environments and
recording devices.

Note that there is no expectation to produce perfect results that match the
quality of an audio recording studio, I'm more in the "rough, but good
enough for practical purposes" territory.

Having read the documentation and various forums, I put together this
pipeline (actual commands in the appendix):

1. run volumedetect to see what the maximum level is
1a. parse stdout to extract `max_volume`
2. normalize audio to `max_volume`
3. apply silenceremove with <empirically determined threshold>
3a. for the beginning of the file
3b. invert the stream and run another silenceremove for the beginning
(which is actually the end)
3c. invert it back and save the output



What I read in the forums gave me the impression that we need step#2 such
that at step#3 we could say the threshold is 0. However, that is not the
case, I still had to find a reasonable threshold via trial and error.

After I found a value that produces a good result, I assumed that it might
be good enough for practical purposes and it would be OK to simply hardcode
it into my code as a magic number. However, on the next day I attempted to
replicate the results using the same recording device in the same room -
but this time ffmpeg would tell me the filtered stream is empty, nothing to
write. The environment wasn't 100% identical, since I'm not doing this in a
controlled lab, but most of the variables are the same, though perhaps the
windows were open and it was a different time of the day, so the baseline
noise level outside was somewhat different.

Clearly, my approach is not robust. I'd like to understand whether there
are any low-hanging fruits that I can try, or if I'm not on the right track.

I imagine that the solution I need would somehow determine the silence
threshold relative to the rest of the file, instead of using a "one fits
all" value. However I did not find such filters or analyzers in ffmpeg.


Your guidance will be greatly appreciated,
Alex




Appendix, pipeline commands

1. ffmpeg -i input.mp3 -af "volumedetect"  -f null /dev/null
here I parse stdout, looking for something like "[Parsed_volumedetect_0 @
0x559dbe815f00] max_volume: -15.9 dB"

2. ffmpeg -i input.mp3 -af "volume=15.9dB" out2-normalized.mp3

3. ffmpeg -i out2-normalized.mp3 -af
silenceremove=start_periods=1:start_duration=0:start_threshold=-6dB:start_silence=0.5,areverse,silenceremove=start_periods=1:start_duration=0:start_threshold=-6dB:start_silence=0.5,afade=t=in:st=0:d=0.3,areverse,afade=t=in:st=0:d=0.3
out3-trimmed.mp3


An example of an input file is available at
railean.net/files/public-temp/in-fresh.mp3, after normalization you can
hear some church bells in the distance. I'm totally fine with them
remaining audible in the result, as long as the leading and trailing
silence is removed.