In an interview with EE Times, Paul Beckmann, CTO and founder at DSP Concepts, highlighted how the Tensilica HiFi 5 DSP includes hardware acceleration to enable voice-controlled user interfaces and artificial intelligence (AI)-based audio applications.

“The complexity of audio products continues to increase especially as more product makers want to add voice and AI features,” said Beckmann.  He added, “the HiFi-5 processor is ideal for these applications and has been designed with high computational throughput and low power consumption.  The HiFi-5 will enable the next generation of smart wearables devices with contextual awareness, a built-in voice assistant, and long battery life.”

Audio Processing
Digital signal processing, whatever its nature, is undoubtedly one of the technologies that has received the most success and attention in recent years, with multiple applications in various areas of electronics and applied sciences in general. Audio information at the same time is assuming an increasingly important role in the context of both industrial and consumer applications, with various technologies being developed. Poor audio processing is the death of productivity in audio conferencing. Hearing details without unwanted noise, echoes or distortion, especially in group conference calls, is crucial and involves many technologies.

“Each audio product is different and you have to carefully consider what algorithms are needed including both microphone input for voice and speaker playback for output. If the product has a loudspeaker, then you will need an acoustic echo canceller (AEC). If it has multiple speakers, like a soundbar, then a multichannel AEC is required. A microphone array improves the performance of the product in noisy environments and with reverberation,” said Beckmann.

“Just like Tensorflow is used for machine learning development, Audio Weaver is used for audio product development,” said Beckmann.  “Audio Weaver contains all of the lower-level pieces of IP (beamformers, echo cancellers, noise reduction, speaker equalization, etc.) and product makers can use Audio Weaver to quickly and efficiently engineer their products.”

Digital signal processors represent a computing architecture for the implementation of audio/visual processes that by their nature refer to signal information and therefore require intensive computational processing activities.

Field programmable gate arrays (FPGAs) and MCUs represent another computing solution thanks to the possibility of implementing audio/visual signal processing algorithms with programmable hardware. The alternative between these solutions depends on a common requirement of audio/visual applications, real-time processing. The real-time requirement is particularly stringent for audio (but also visual) applications since, especially for the visual component, the computational level is particularly high.

“MCUs, DSPs, and application processors can all be used for audio processing. Real-time means that the audio workload must be able to complete within a fixed time interval.  This requires careful profiling of the processor to ensure that the workload will fit. This requires accurate profiling tools which operate in real-time and can pinpoint hot spots where significant processing occurs,” said Beckmann.

He added, “DSPs have advantages relative to MCUs and application processors when it comes to power consumption. The DSP architecture is optimized for mathematical operations and has a high degree of parallelism. We have found that the HiFi-5 DSP requires the lowest number of instructions to complete commonly used DSP functions like filters and echo cancellers.  A more efficient processor architecture translates to lower power consumption.”

The main challenge for DSPs is that they are harder to program and optimize for.  However with Audio Weaver, Beckmann highlighted that DSP Concepts has handled the low-level programming and optimization allowing engineers to design and tune algorithms graphically.

 

Figure 1: DSP Cadence Tensilica HiFi 5. Click to enlarge the image.

DSP Hardware and software
The exponential growth in the adoption of voice interactions for a wide range of devices has spurred the need for increased signal processing and neural network performance. To do this, Cadence is working with DSP Concepts to enable processing algorithms on the Tensilica HiFi 5 DSP to improve full-duplex communication, voice user interfaces performance and playback processing (figures 1 and 2).

Figure 2: Audio Weaver Designer

Cadence points out that this fifth-generation HiFi DSP offers four times the processing performance of the HiFi 4 DSP, making it an ideal solution for voice-controlled user interfaces in digital home assistants and automotive infotainment. The HiFi 5 DSP offers improved fixed and floating-point DSP capabilities and native support for new data types, saving memory and power.

Due to the requirements for lower latency, greater privacy and more natural voice UI interactions demanded by consumers, the processing load on the device is increasing rapidly. The HiFi 5 DSP with Audio Weaver provides the performance needed to handle the front-end processing tasks, including echo cancellation and noise reduction.

At the heart of these software and hardware innovations are intelligent turnkey solutions that include far-field voice UI for smart assistants, TVs, soundbars and set-top boxes. These solutions are also optimized for enhanced voice communication and video conferencing at home and in the office, supporting 2-microphone, 4-microphone, 6-microphone and customized configurations with acoustic echo cancellation (AEC) for up to 7 channels.

“Both wearable devices and conferencing devices can leverage AI for audio. In both applications, AI would be used to eliminate noise and focus on the main speaker. AI can be used to authenticate the owner of the wearable device. More sophisticated AI algorithms can be used in the conference room application because it has a larger processor and there are no battery constraints. The upcoming HiFi 5 processor from Cadence will allow complex AI applications even in low power. This will expand AI applications in wearable devices, especially truly wireless systems which is very power sensitive,” said Beckmann.

Combined with optimized Audio Weaver libraries, Cadence’s Tensilica HiFi 5 DSP will have access to the full suite of DSP Concepts’ audio front-end (AFE) TalkTo which provides multi-channel echo cancellation solutions with advanced beamforming using 8 or more microphones. Additionally, the new machine learning (ML) instruction sets in the Tensilica HiFi 5 DSP enhance full duplex communication and playback processing.

Audio Weaver aims to provide clear voice communication in the most challenging environments. Beckmann pointed out that the library not only offers the processing required for a correct voice user interface but also supports the power performance for all battery-powered solutions.

Audio Weaver allows designers to create complex digital audio applications without writing a single line of code. Developers simply choose the necessary audio modules from a library of 500+ modules, link them graphically, and adapt them to the target hardware in an equally simple process; the included examples accelerate learning and allow new designs to be implemented more quickly.

The speaker has highlighted that DSP Concepts has a range of rapid prototyping hardware which can be used early on for development. This hardware can be used early on to answer performance questions and de-risk product development.