The conversation of one or more people will be analyzed. I think need to look for something similar to reverberation or echo in an audio signal. How can this be implemented?
Update: small room is about 10-15m2, large room 50-100m2
4
-
1$\begingroup$ What is large and what is small? A small room could be a 10 cm cubic room as a part of a dollhouse, a cubic 1.5-meter prison cell or a room too small to occupy two children in a house. Can you specify the parameters of the problem? I think you can train a NN for such a purpose given a good dataset. Otherwise, estimate the RIR and analize the reflections. $\endgroup$– havakokCommented Jan 14, 2020 at 14:01
-
$\begingroup$ Do you have the ability to use a known sounding signal? $\endgroup$– Dan BoschenCommented Jan 14, 2020 at 16:56
-
$\begingroup$ Nope, it's just people talking or just sounds in a room $\endgroup$– eugennCommented Jan 14, 2020 at 20:08
-
1$\begingroup$ Not if they're both anechoic... $\endgroup$– endolithCommented Nov 9, 2020 at 17:16
Add a comment
|
3 Answers
$\begingroup$
$\endgroup$
2
There could be a number of approaches for this task, depending on the number of channels (microphones) in your input and structure of available data.
Given a labeled data set from a previously known set of rooms, you can train a neural network (NN) to classify in which one of them a new sample was recorded.
A second approach would be to estimate the parameters of the room impulse response (RIR). Certain methods could be under the limitation of a single microphone. A microphone array would supply more information like in this paper. Depending on the shape of the array, different methods can be utilized.
Also, do you have any other prior knowledge on the room? For example, can you get a clean (without speech) audio segment? Do you know where is the microphone and sound sources are located on the room? These priors can point you to methods that utilize this information and improve the estimation.
-
$\begingroup$ I would like to try using several microphones built into the laptop. the sound sources can be different and located in different places of the room. $\endgroup$– eugennCommented Jan 14, 2020 at 20:18
-
$\begingroup$ The laptop contains several built-in mics?!? What kind of laptop is that? $\endgroup$– BrianOCommented Mar 6, 2022 at 19:19
$\begingroup$
$\endgroup$
Dry speech contains somewhat steep and deep «tails» at the end of each utterance if you look at a (magnitude) spectrogram.
Long room reverberation will tend to smear this decay out, and (along with any noise) tend to reduce the magnitude spectrogram «contrast».
-k
$\begingroup$
$\endgroup$
If you were to plot the log of the absolute value of the audio data, then in the spaces between words and sentences, I would expect to see roughly exponential decay, which will appear as falling straight lines bounding from above a very ragged signal (because some data values will be quite small, and have a large negative log.)
The rate of this decay (the slope of the falling lines), in dB/sec, will be slower in a larger room. I can't supply any numbers, but if you have samples of voice recordings in both large and small rooms, my guess is the difference will be clear. Automating this procedure would be challenging; it's not clear if that is a requirement.
