As many have probably experienced it first hand, the echo suppressor (ES) in pjmedia sucks, to put it mildly. And this has made matters worse since ES plays such a major role in PDA’s and/or mobile devices/smartphones, since echoes are heavily present on these devices and we can’t really put the Accoustic Echo Canceller (AEC) on these devices due to the high processing requirements of the AEC (see our results in Evaluating PJMEDIA Performance article).
So we scrapped the old echo suppressor and wrote another one from scratch. And I think this one works quite well.
First of all, it’s low-complexity so it doesn’t require high processing requirements. We tested on a PocketPC 2003 PDA with 312MHz StrongARM CPU, it only uses less than 0.5% (half percent) of CPU. So it’s very affordable for these devices.
And most importantly, I think it works! It suppresses echo significantly, it lets non-echo audio pass through, and also it allows double-talk where both parties are talking at the same time. This is a very important to allow enjoyable conversation between two people, so we’ve made sure that the new echo suppressor supports this.
For those who care with the details, the ES works by comparing the audio level pattern in the input signal from the microphone with the audio level pattern in the playback signal during the learning process. It calculates the correlation values for each tail position up to the configured echo tail length, and finds out which tail position has the best correlation value. This position is then marked as the echo position. While doing this, the ES also notes the gain factors, that is how much de-amplification to be applied to the input signal in order to remove the echo from the input. Once the tail position is found, the ES then applies the appropriate gain factor according to the state of the conversation (e.g. I’m talking you’re silent, you’re silent I’m talking, we’re both talking, etc.).
This works pretty well. On the downside though, we noticed that the ES still seems to let a quiet echo to pass and transmitted back to remote party, depending on the signal level at the time. For now we allow this to happen since the echo level is very low and doesn’t seem to be too annoying, and lacking enough experience with the new algorithm we don’t want to make it cut signals too aggressively yet since it may inadvertently cut the “good” signal.
The new echo suppressor is available in SVN as pjmedia/echo_suppress.c, and it will be included in the next release.