Hello,
I have been working on acoustic echo cancellation while following the research paper :
Conference Paper Echo Detection and Delay Estimation using a Pattern Recognti...
Since I am working on real-time audio data so my problem is as follows:
- I have a buffer that stores Far-end (packets being played by phone) data in terms of 21.33ms chunk equivalent to 1024 shorts.( Sampling rate 48000 Hz)
- A near-end packet is recorded i.e 21.33ms of data with the same format as mentioned above.
Problem Statement:
- Let's suppose we have a Far-end packet containing the word "hello"
------------------------------
| |
| H E L L O |
------------------------------
is being played by phone and now its echo is being recorded in a Near-End packet. Now the cases arise here are that this could be recorded as:1
------------------------------
| H E L L O |
------------------------------
Hello completely recorded in one packet.
2. ------------------------------ ------------------------------
| | | |
| H E | | L L O |
------------------------------ ------------------------------
Distributed in two Near-end packets
3. ------------------------------ ------------------------------
| | | |
| H | | E L L O |
------------------------------ ------------------------------
and so on random distribution of far-end audio between multiple chunks of near-end audio.
Now, I want to detect echo and perform cancellation. That can be done if I get an accurate chunk of far-end whose echo is in the near end chunk. Right now I am making overlapped chunks with the shift of 1 sample equivalent to 0.020833 ms of data. But this is quite computationally over expensive.
So 2 questions arise here:
I hope my question will be cleared. I tried to explain it yet keep it concise.
Any help will be appreciated.
Regards,
Khubaib Ahmad
if you want to process audible echoes, the chunks needs to be no less than 80-100ms long. Audible echo limit in speech is about 50 ms.
Anders Buen thank you very much for the specific detail. But I am looking for a programmatic answer and let me add some pictorial representation of my question. It's messed up above.
So if you can see the pictorial representation (sorry for the messed up drawing), my question is the optimal step size here that what step size should I choose that will not affect the accurate detection of echo meanwhile decreasing the overall computational expense.
You got to learn how the brain and listening system processes sound and echoes in particular, to get this to work. If you do not understand this you, will never be able to develop system that works!
Anders Buen so according to your answer should I consider echoes only greater than 50ms?
For the part relating to the listeners, yes. For the recognition and measurement an earlier echo detection may be useful for potential risky ones. I have never processed this myself, but have a lot of room acoustics measurement and prediction experience. To figure out if there are echoes in rooms or measured impulse responses listening to the impulse responses is the best method in my experience. You need to take that further. Althoiugh video conference systems and possibly mobile phones have such algoritmes already. Theree should be vast amounts of literature on the phenomenon.
I also have a conference article on the acoustics of coupled rooms. The environments needs to be drier than usual to work well. It is impossible to get rid of reverb that are already on the signal..
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python