Use auditok as API to detect pauses in speech

[bwang482]

Thanks for the great lib!!

I have a bunch of utterances extracted from conversations, and I want to detect pauses in each of these utterances and how long the pauses are, both short and long pauses (500ms is the threshold for example for determining long or short).

Is it possible to use auditok as API to call such function for pause detection, in my existing data pipeline at all? Sorry if the question seems general, but it'd be greatly appreciated if you can provide any advices. Thank you again.

[amsehili]

Hello,

Audio regions returned by the split function (or by AudioRegion.split or AudioRegion.split_and_plot methods) contain the start and end information in their metadata. You can extract pauses' onset/offset which are simply the end of region i and the start of region i+1. You also need to deal with corner cases which are:

• 0 and the start of the first region if the first region does not start at 0
• The end of last region and the duration of the whole file if the last region ends before the end of the file

So concretely you can do something like this:

from auditok import AudioRegion
from auditok.dataset import one_to_six_arabic_16000_mono_bc_noise
region = AudioRegion.load(one_to_six_arabic_16000_mono_bc_noise)
regions = region.split_and_plot(energy_threshold=65, drop_trailing_silence=True)

# print regions' start/end
for r in regions:
    print(r.meta.start, r.meta.end)

# extract pauses
onset = 0
pauses = []
for r in regions:
   if onset < r.meta.start:       
       pauses.append((onset, r.meta.start))
       onset = r.meta.end

# and the last pause if there exists
if onset < region.duration:
   pauses.append((onset, region.duration))

Remarks

• If you don't have matplotlib installed, just call list(region.split(...)) instead of split_and_plot(...)
• I'm using an energy threshold of 65, this might not be what you need for your data, so you need to either plot detected regions (using split_and_plot, hence installing matplotlib) or listen to them (requires pyaudio) in order to choose an appropriate value for this parameter.
• I'm using drop_trailing_silence=True which is not recommended when you want to detect/use speech but since you're interested in detecting pauses, setting this parameter to True would eventually yield more accurate pauses' start end end.
• You should explore other parameters of split, especially the ones dealing with durations (https://github.com/amsehili/auditok/blob/master/auditok/core.py#L42).