Icelandic audio startup Treble Technologies ehf today announced that it has raised $18 million in funding.
Paladin Capital Group led the Series-A2 round. It was joined by returning backers KOMPAS VC, Frumtak Ventures and the European Innovation Council Fund.
Many industrial robots rely on onboard artificial intelligence models to perform their work. Those AI models often collect the data they use to make decisions with microphones. For example, a robotic arm might process spoken instructions from human factory staff. Autonomous vehicles listen for distant collisions and construction noises.
Engineers teach robots to process audio by giving them examples of the sounds they will be expected to analyze. Those training examples historically had to be recorded manually, which is a highly time-consuming process. Treble offers a platform that speeds up the workflow.
The company’s software simulates the environments in which a robot is designed to operate. From there, it generates synthetic sounds similar to the audio that the robot will be expected to process. The Treble-generated synthetic audio teaches the machine’s onboard AI how to interpret acoustic data.
Engineers have to take numerous factors into account when generating audio datasets. One is that the size of an indoor space influences how sound travels through it. The materials from which the facility is constructed, its background noise and the configuration of a robot’s microphones also factor into audio quality. Plus, sounds that originate from moving objects are often more difficult to process than a static audio source.
The company says its simulations can account for all those factors. The company’s platform enables users to set up a simulation by uploading a blueprint of the indoor space they wish to replicate. From there, they can use Python code to customize the virtual environment. That’s significantly faster than physically recreating every single combination of acoustic conditions a robot is expected to navigate.
Treble outputs the audio that it generates in the form of labeled datasets. That means the individual sound snippets are paired with explanatory metadata, which eases AI training. Furthermore, the software can automatically run parameter sweeps. Those are workflows that improve an AI model’s accuracy by optimizing its configuration settings.
The startup offers its platform alongside pre-packaged audio datasets that remove the need for developers to set up simulations. According to the company, its datasets can reduce the error rate of speech recognition models by 38%. Treble says that they also lend themselves to other use cases.
“AI models and devices that perform well in a laboratory can struggle when they encounter reverberation, background noise, or an unfamiliar physical environment,” said co-founder and Chief Executive Finnur Pind. “Treble gives development teams at the world’s leading device makers a faster and more scalable way to model those conditions, generate the data they need, and test products to ensure they work reliably.”
Treble will use the proceeds from its funding round to enhance its platform and accelerate go-to-market initiatives.
