Call centers in the United States generate more than a billion hours of phone audio every year, at a cost of over $10 billion. Despite this huge investment, less than 25% of this audio is currently made searchable or analyzed. Other industries, such as legal discovery and sales, also have large amounts of neglected but valuable audio content.
DeepGram is a company launching out of our Winter 2016 class that uses deep learning to index audio and make it searchable for businesses. For example, a company can use DeepGram to analyze their phone support audio dataset and search for moments where their competitors' names are mentioned.
DeepGram has been made possible thanks to the convergence of several relatively recent technical developments: Cloud storage has enabled companies to store their recorded audio cheaply; and advanced GPUs have allowed deep learning to mature, so that it's now possible to build deep neural networks that can capture the complexity of speech.
Speech search is also motivated by market factors. There has been a structural change in phone support from on-site employees to a distributed workforce which makes quality assurance more challenging. Businesses are focusing more on data and analytics, and they want actionable insight from their information-rich audio datasets.
status quo approach to the audio search problem is to use speech
recognition to create a transcript of the audio, and then use a
conventional text search to search the transcript. But the inevitable
errors in the computerized transcript make this work poorly. DeepGram is
far more accurate and efficient: For example, in the image above, DeepGram correctly identifies the word "turbine" where the speech transcript misreads it as "turban."
DeepGram was founded by two particle physicists searching for dark matter. The idea for the audio search engine came from a weekend project of theirs, to record all the audio in a person’s life 24/7. But with so much audio data, it was impossible to find and recall certain moments, and they couldn’t find a search engine that worked well with speech. So, they built an AI-based index and fuzzy search, making their recorded audio accurately and easily searchable. They realized the search engine would be even more useful for businesses with large amounts of audio -- and DeepGram was born.
Businesses can sign up and start using DeepGram now at www.deepgram.com.