Audio and video technologies, have advanced rapidly over the last decade. Narrowband (8kHz sampling) audio was the most common bandwidth for audio. With the increased importance of audio conferencing for remote work environments, it is desirable to achieve the best audio quality possible. The transition from PSTNs to high-speed mobile wireless networks (4G, LTE, and 5G) has made tele-work possible. The evolution of the audio equipment in mobile devices and their networks allow for the utilization of wideband (16kHz sampling) audio to fullband (48kHz sampling) audio.
Despite low-cost embedded devices having more CPU speed than ever before, engineers are always looking to minimize power consumption while looking to add features that would make their product standout. One of the biggest resource consumers in the full-duplex audio communication realm is acoustic echo cancellation (AEC). AEC require long adaptive filters, and increasing the sampling rate of audio causes a double hit complexity. The complexity not only increases because you have to process more samples per second, but because the filter length has to increase to achieve the same tail length in time.
The table below shows number of multiplications per second (MPS) required for a time domain FIR and NLMS filter update. As it can be seen observed, to increase the sampling rate from 8kHz to 48kHz the number of MPS increases 36-fold.
| Sampling Rate | Tail Length | FIR Multiplies Per Second | NLMS Update Multiplies Per Second |
|---|---|---|---|
| 8 kHz | 128ms | 8,192,000 | 16,416,000 |
| 8 kHz | 256ms | 16,384,000 | 32,800,000 |
| 16 kHz | 128ms | 32,768,000 | 65,600,000 |
| 16 kHz | 256ms | 65,536,000 | 131,136,000 |
| 48 kHz | 128ms | 294,912,000 | 590,016,000 |
| 48 kHz | 256ms | 589,824,000 | 1,179,840,000 |
Time-domain multiplies for AEC
All hope is not lost. As discussed in Subband Adaptive Filtering, the computational resources can be significantly reduced by transforming from the time domain to the subband domain to perform the AEC filtering. The table below shows that the MPS for 128ms tail length for various sampling rates and subband decimation factors.
| Sampling Rate | # of Subbands | FIR Multiplies Per Second | NLMS Update Multiplies Per Second |
|---|---|---|---|
| 8 kHz | 32 | 1,056,000 | 2,116,125 |
| 16 kHz | 64 | 2,080,000 | 4,164,063 |
| 48 kHz | 128 | 9,288,000 | 18,582,047 |
| 48 kHz | 256 | 4,626,000 | 9,255,012 |
Subband domain multiplies for AEC
For 48kHz with 256 subbands, the MPS is reduced nearly 64 times over compared to the time domain implementation. There is an overhead to transform the data to and from the subband domain, but FFT’s can be heavily optimized and some DSPs have hardware FFT accelerators. In addition to the subband optimization, higher sampling rates can take advantage of the fact that higher frequencies are more easily absorb and therefore require shorter echo tails at higher frequencies. This characteristic of acoustics critically important for professional audio-conferencing applications where high-resolution audio (96kHz sampling) is used.
At VOCAL Technologies, our wideband, super-wideband, fullband, high-resolution AEC is efficiently implemented in the subband domain and optimized for a wide variety of embedded DSPs. Please contact us to learn more.