With the rise in popularity of voice UI systems has come a rise in reports of system error. Reports of a smartspeaker accidentally making a phone call or sending an e-mail appear in the media often, putting pressure on manufacturers to make these systems better at distinguishing actual commands from false triggers.
DSP Concepts has found that the best way to assure accurate recognition of voice commands is to increase the level, or volume, of the voice command relative to the level of surrounding noise and the level of the program material coming from the device the voice UI system is built into. With voice command systems becoming more common in automotive and portable use—both relatively loud environments—the challenge of separating voice commands from noise and interference becomes even greater.
This is why DSP Concepts has developed its exclusive Adaptive Interference Canceller, or AIC. AIC uses machine-learning techniques to work differently—and better—than other noise cancelling or noise reduction technologies on the market.
Most voice UI systems combine two basic techniques to improve the ratio of the voice signal to the surrounding noise. One is acoustic echo cancellation (AEC), which attempts to remove any recorded sound that originated from the devices speaker(s). To do this, AEC algorithms look for a reference signal (the audio fed to the speaker) in the mic recording, and then attempts to cancel it by adding it's inverse. The other is a stationary noise reduction algorithm that uses classical signal processing techniques to reduce noise that is relatively consistent, such as road noise and wind noise. These techniques work—in fact, DSP Concepts uses them, too—but they are inadequate to deliver the nearly noise-free voice signals that voice command systems need to function reliably when there are competing interference sounds.
Adaptive Interference Cancellation uses machine learning techniques to reduce noise in much the same way a professional recording engineer would deal with a noisy recording. It continuously listens to the environment to identify the source and character of the most prominent noises, then cancels them out automatically—even if they are inconsistent and unpredictable, such as music and speech. AIC works so well that it reduces the level of noise by as much as 30 dB.
Because AIC re-evaluates environmental noise continously, it adapts almost instantly to any new environment. This makes it ideal for portable applications, as well as for automotive applications, where the level and character of the noise is constantly changing due to speed fluctuations and differing road conditions. It works with as few as two microphones, so it can be incorporated into tiny wearable devices and low-cost voice UI products.
AIC also overcomes one of the most troublesome limitations of AEC: inability to process more than two channels of sound. As more channels are added to a playback system, the processing load on an AEC system increases by a factor even greater than the increase in channels. Trying to do AEC on more than two channels overwhelms the DSPs and SoCs used for voice UI systems. This might not be an issue in the home, where there might be only one or two major sources of interference, but in automotive systems, which commonly employ a dozen or more channels, it can greatly degrade the response and accuracy of a voice-recognition system.
For AIC, the number of playback channels is irrelevant—because like a recording engineer, it listens for the most troublesome sources of noise and interference, and eliminates them. AIC doesn’t care if the interfering sound is road or mechanical noise from the auto, or sound from the car’s audio system. It simply identifies the interference and eliminates it.
Combined with DSP Concepts’ noise reduction and AEC algorithms, AIC can reduce noise to the point where a voice UI system will noticeably outperform the systems currently on the market. Tests show that a DSP Concepts-powered voice UI system can reliably process voice commands from more than twice the distance of the Amazon Echo Dot, which is currently the comparison standard for the tech industry. In fact, customers have found that DSP Concepts AIC achieved sufficient noise and interference rejection that they didn’t need to use AEC at all in automotive applications. This greatly reduced the processing demand and allowed them additional flexibility in system design.
AIC is included as a module in DSP Concepts voice UI and automotive algorithms, and in Audio Weaver®, our modular DSP programming environment. Thus, it’s easy for any DSP Concepts customer to add AIC to any voice UI product. At no extra cost, the product works better—with fewer worries that it might inadvertently place an embarrassing phone call to the customer’s boss.