Signal Processing - Echo Cancellation

Method 1 - Adaptive filters

1.LMS method

To figure out W(n), the filter we use least mean squared method. (I have developed LMS algorithm before)

To figure out time delay of the echo we use cross correlation (The mathematical method which I explained in the meeting)

Difficulty - Medium

Effectiveness - Good

Computational complexity - Medium

Implementation - 1-2 weeks.

2. Adaptive FIR filter

To figure out W(n), the filter we use Finite Impulse Response Filter method. (I have developed FIR algorithm before)

To figure out time delay of the echo we use cross correlation (The mathematical method which I explained in the meeting)

Difficulty - Medium

Effectiveness - Good

Computational complexity - High

Implementation - 1-2 weeks.

2. Noise Injecting Method

3. Kalman Filter Method

4. Neural Network (U-net)

5. ICA

This is algorithm is commonly used to unmix a signal where few speakers speak, in to single speaker signal. See cocktail party problem. You can see my precious implementation here.

Difficulty - Easy

Effectiveness - Medium

Computational complexity - High

Implementation - 1-2 weeks

6. Conclusion

I think if we first approach an adaptive filter method that would be great because that's the most popular method. In-fact other methods are built on top of that . Same concept. we can easily swap to another method.

As the second best I would recommend second approach (Noise injecting algorithm)

AI solution implementation is easy but has to do lot of training which requires time.

7. Regarding Working Hours

Estimated work hours per week - 36 hours ( My current project in the company will be finishing coming week. So until I apply for other jobs I may have 2-3 weeks left to do this project. In that case I can work 8 hours per day on this project. So 40 hours per week , not including weekends)