Follow Datanami:
June 4, 2020

In-Q-Tel Enlists Deepgram for Speech Recognition

Staff report

Deepgram, the speech recognition vendor, has signed on with the investment arm of the U.S. intelligence community.

The Bay Area startup said Thursday (June 4) is has inked a strategic investment agreement with In-Q-Tel. The agreement allows government agencies to use Deepgram’s deep learning platform to improve the accuracy of speech recognition and transcription tools.

The company’s platform scans audio data to train a speech recognition tool. Deepgram’s deep learning approach uses a hybrid CNN (convolutional neural network)/RNN (recurrent neural network), with models trained on GPU accelerators.

The result is a substantial improvement in transcription quality along with the ability to translate speech data in a variety of languages, the company said. “It’s optimized to deliver better accuracy under real-world conditions,” Deepgram CEO Scott Stephenson told sister web site EnterpriseAI last month.

“Deepgram’s use of an AI-enabled, neural network architecture leveraging custom speech recognition models trained on vast amounts of audio data allows them to rapidly achieve much more accurate transcriptions for non-standard audio environments vs solutions like Google Voice and Apple Siri,” said George Hoyem, In-Q-Tel’s managing partner for investments.

“Using state of the art transfer learning also enables Deepgram to quickly build speech-to-text capabilities for new and novel language variants on relatively small amounts of training data, resulting in huge time savings for our government partners,” Hoyem added.

Prior to it agreement with In-Q-Tel, Deepgram raised $13.9 million in venture funding.

Recent items:

Speech Analytics Makes Unexpected Discovery

Voice: The New Interface for Analytics

Intel’s Sentiment Analysis Algorithm Needs Less Training