Done! A bit impressed with myself.

Our product needs the function of echo cancellation, three possible technical solutions were identified,
1) use MCU to detect audio out and audio in audio signals, write an algorithm to calculate the strength of the two sides of the sound signal, according to the strength of the audio out and audio in between the two optional channel switching, to achieve the half-duplex call effect, but now on the market are full-duplex call effect, half-duplex will make the product lower competitive

(2) Use the echo cancellation algorithm from CPU vendor, the actual test echo elimination is not clean enough by all the adjustable parameters tuning and many round discussion with vendor, far from the expected effect. the feedback from vendor is to improve the housing to isolate the MIC as far as possible from speaker, But no room to change the housing due to ID design, sound quality and volume requirements.

(3) Download the open source echo cancellation algorithms such as webrtc and speex from the Internet, and port them to the product after cross-compiling.

4) Purchase the algorithms from a company that specializes in audio algorithms, but each product would require additional costs, which would greatly reduce the competitiveness of the product.

After weighing the options, I decided to do an in-depth research on the solution of using open source algorithms;

So, I downloaded a series of echo cancellation codes from github, gitee and other sites, both in C and matlab.

Synthesize the two voices with octave to generate a section of a double-talk scene with the near-end mic capture voice as well as the far-end reference voice.
Then run the code downloaded online for echo cancellation, analyze the output audio and pick the algorithm that works the best among them.

From the results, the AEC algorithm of webrtc audio processing is not good, the echo cancellation is not clean, and there is obvious swallowing of words in the double-talk scene. The AEC3 algorithm of webrtc audio processing is clean, but it suppresses the sound of the other end of the double-talk very much, and the sound is intermittent and unnatural;

The speex has a slight echo, it is best to find a matlab language implementation of the aec algorithm, echo elimination is very clean, double speak only a slight swallowing of words phenomenon.

Then, I use visual studio to compile and debug the open source code of webrtc audioprocessing aec, refer to the above aec algorithm of matlab to modify the code, use online debugging to set breakpoints, single-step running and other analysis of the value of the statement-by-step changes, and ultimately get the data and the results of the aec algorithm of the matlab language are identical The final data obtained is exactly the same as the result of the aec algorithm in matlab;

Today compiled to the product, the MIC and SPEAKER gain adjusted to a reasonable value, and then the actual call to test the effect of echo cancellation, the results are surprisingly good.

Thereby, there is hope to design a product that can reach sales above the small target.