phonlamaiphoto - stock.adobe.com
Sharpen, a vendor of a cloud-based contact center platform, offers an automated transcription service as part of its software package. It's free, which its customers like, but a few years ago, Sharpen was getting complaints from customers that its transcriptions weren't accurate.
"The transcription wasn't great," said Adam Settle, Sharpen's vice president of product, who declined to name the automated transcription software vendor that the company used.
The customer wanted to use the service for keyword spotting, he said. But, he added, "searching for a keyword is kind of pointless" if the transcription is wrong.
The complaint sparked Sharpen to search for a new automated transcription vendor. That search eventually led them to Deepgram, an automated speech recognition startup founded in 2015.
A new vendor
Sharpen first became acquainted with Deepgram a few years ago, after seeing it demonstrate its automated speech recognition platform at a conference.
The platform, built on deep learning models, can come pre-trained on Deepgram's library of calls. Users can upload pre-labeled speech files or label speech as they go, to further train and customize the platform.
Users can run the platform in the cloud or on premises and can access the speech recognition models through APIs.
Sharpen tested the products of numerous startups and big-name tech vendors before choosing Deepgram. Each had its problems, however. Some platforms, like the one from their first vendor, didn't provide accurate transcriptions. Others, including systems from Google and Amazon, were too expensive, Settle said.
Some platforms "were eight times the cost without being eight times the quality," he said.
Sharpen, after all, offers the transcription service for free to its clients. Deepgram, charging about 23 cents per hour, is about half the platform's cost from Sharpen's previous vendor.
Settle noted that Sharpen created a general transcription model, dubbed the Sharpen model, in Deepgram that all but one of its customers use.
The model, which took less than a week to build, can pick up on everyday speech, as well as common jargon words Sharpen's customers use. Sharpen built a custom model for one of its biggest clients, the same client that complained about the previous automated transcription vendor.
That client, which needed a custom model due to its heavy use of jargon, isn't complaining anymore, Settle said. The company uses Deepgram's automated speech recognition heavily, transcribing hundreds of hours' worth of calls each month.
Using automated speech recognition
Almost all of Sharpen's customers use the automated transcription service, Settle said. Since it's free to them, many use it just because they can, on the off chance the transcription will benefit them later on. Many also rely on the platform for targeted uses, including identifying robocalls by directing the platform to pick out certain words or phrases commonly used by robocallers and coaching sales staff.
Adam SettleVice president of product, Sharpen
Settle said he doesn't think Sharpen has ever had a problem with Deepgram and hasn't asked the vendor for technical support outside of the initial setup. The product runs without much supervision, allowing Settle and other Sharpen employees to focus on other tasks.
"It's super nice to not have to think about it," he said.
Deepgram currently supports about ten languages and has recently begun to support real-time transcription. Settle said Sharpen plans to incorporate both features into its platform, once it works out an interface for it on their end.
Multi-language support, in particular, will help Sharpen scale out to other countries and populations, Settle said, potentially enabling the vendor to grow faster.
Deepgram recently raised $25 million in Series B funding led by Tiger Global, with Citi Ventures, Wing VC, SAP.io and Nvidia Inception GPU Ventures. The startup announced the funding Feb. 3.