Chip simplifies integration of voice detection algorithms for cloud service

Article By : Majeed Ahmad

A new chip for voice-enabled designs offers an API that simplifies the integration of voice capture algorithms into different cloud services.

While designing smart speakers and other voice-enabled devices such as wearables and hearables, developers’ primary challenge is to make microphones accurately and efficiently detect the wake words like “Alexa” and “Hey Google” for processing in the cloud.

When DSP Group launched its voice-enabled system-on-chip (SoC) DBM10, EDN brought forward this issue to Yosi Brosh, product manager for the company’s SmartVoice chips. The dual-core SoC—based on a DSP and a neural network (NN) accelerator—is optimized for voice and sensor processing in battery-operated devices such as hearables, wearables, true wireless stereo (TWS) headsets, and smart home remote controls.

DBM10 chip block diagramFigure 1 The DBM10 chip for voice-enabled smart products claims to have a platform approach with comprehensive software framework support. Source: DSP Group

Brosh said that a cloud platform like Amazon Web Services (AWS) is not interested in configuring a register on a chip. Instead, these cloud services are focused on how voice algorithms can efficiently detect wake words. “They want algorithms to detect wake words without engineers spending a lot of time on studying and configuring registers on a chip,” Brosh said.

Therefore, DSP Group has developed an API that makes voice detection algorithms easy to integrate into cloud services. “In a way, voice algorithms from cloud service providers become a black box for the chip,” he added.

The common practice has been that device makers release software code with microphone settings and tell algorithm providers how to integrate algorithms in microphone drivers. In the case of DSP Group’s DBM10 chip, it captures the audio using voice firmware and makes the integration of voice capture algorithms efficient and simple.

DSP Group has been closely working with more than a dozen cloud firms—including Alibaba, Amazon, Baidu, Google, and Samsung—while porting their voice algorithms on its silicon. According to Brosh, the company are also providing a complete software suite in some cases.

As to the chip’s ability to provide an easy deployment path for system designers, Brosh said that the company’s support for software running on the DBM10 voice interface chip is provided all the way to the production level; “System engineers don’t need to write a single line of code.”

That’s why DSP Group calls its DBM10 chip a complete solution. The SoC is optimized for audio algorithms as well as sensing AI algorithms with a generic DSP and a neural network processor called nNetLite. Moreover, DSP Group provides additional drivers running on the Wi-Fi chip for communication with the DBM10 chip.

The SoC also features a cross-platform toolchain that supports all commonly used artificial intelligence (AI) and machine learning (ML) frameworks to simplify the algorithm deployment. Engineers can develop, train, and test the algorithm; next, they can save it in a standard format, and the toolchain will take it and create an image for downloading into the SoC.

diagram for the nNetLite compilerFigure 2 The nNetLite compiler allows rapid optimization, pruning, and deploying of any AI/ML model from any framework to the DBM10 SoC. Source: DSP Group

The SoC is available in a tiny form factor of 4 mm2 to get into very small devices like the smartwatch. Likewise, the always-on wake word algorithm running on the SoC’s neural network nNetLite engine merely consumes microwatts.

This article was originally published on EDN.

Majeed Ahmad, Editor-in-Chief of EDN, has covered the electronics design industry for more than two decades.

Related articles:

Leave a comment