Deepgram Unveils Tool for AI to Teach Automatic Speech Recognition to Other AI
Enterprise speech technology startup Deepgram debuted the new Deepgram AutoML training system for automatic speech recognition (ASR) on Thursday. AutoML, a method used to get one artificial intelligence to train others without the need for humans to manually craft and adjust all of the models, has never been applied to ASR before, according to Deepgram, and should make it easier for businesses to incorporate AI into their plans.
Teach to Listen
Deepgram’s platform offers businesses and other enterprise organizations a way to record, transcribe, and analyze audio with end-to-end deep learning. Meetings, customer service calls, and formal presentations are all part of the company’s portfolio. Deepgram claims its automatic speech recognition is far more accurate than the platforms using more traditional models, yet faster and requiring less computing power at the same time. Deepgram’s models are tuneable for the specific industry or company using it, adjusting to the vocabulary and sonic environment where it is placed.
Deepgram already highlights the speed and reduced resource-costs of its platform, the new training tool augments both of those aspects. AutoML is designed to simplify and accelerate ASR setup and analysis while using fewer resources than standard systems. Variations on AutoML are in use already for other technology, including machine vision and natural language processing. It’s a useful way to improve how AI interprets audio and visual input but hasn’t been a part of an ASR platform, according to the company.
The new tool adds another line to Deepgram’s already very full year, starting with closing a $12 million funding round in March. As the ongoing COVID-19 health crisis placed heavy burdens on healthcare providers, the startup donated $1 million worth of its platform’s services to medical care organizations in April. The company brought in another strategic investment in June, this time from the U.S. intelligence community-founded venture capital firm In-Q-Tel, albeit for an undisclosed sum. Soon after that, Deepgram began a new sales approach, providing free, limited to its technology via the new MissionControl system. Potential clients are now able to test out how Deepgram’s technology might be of aid in their projects, with the goal of enticing them to sign up for longer projects.
“When we started, this approach was new, so it was handled internally, and we were increasing accuracy in a way not apparent to the user. Now we are allowing people to [setup and train] models themselves,” Deepgram CEO Scott Stephenson told Voicebot in an interview earlier this year. “If you tried to release this two years ago, people wouldn’t have understood what it was for or what it could do for them. Now, they’re actually becoming more data-driven; they’re not just saying it anymore. Now, if you don’t have data, there’s something wrong.”