Deepgram’s Tailored Speech Models Drastically Improve Sales Analysis

One-size-fits-all speech recognition models miss too much vertical-specific terminology.

The Landscape

Bringing the Patchwork of Sales Enablement Solutions Under One Platform

Some RevOps solutions surface insights. Others give customers the tools to act on them. does it all—so customers can optimize revenue performance in real-time. Converting conversations into an actionable data source is critical for revenue teams. Conversations are where deals get won or lost, but many companies don’t have direct insights into these conversations at scale. helps customers understand what habits of top performers should be scaled across the team. In addition, reps are led by real-time guidance that helps teams grow revenue.

The Challenge

Analyzing Sales Team Conversations from Various Industries

As with all data analysis, garbage in is garbage out. For sales conversations, that means you need to be able to transcribe all the company names, product names, terminology, and specific words so you can analyze when the conversation is about a specific product or company. As Scott Hoch, Head of Data of, said, “You can build the most cutting edge AI model, but if you don’t put good data in, you’re not going to get good results.” Because different industries and companies have different terminology, jargon, product, and company names, found that improving the specific word accuracy of the one-size-fits-all Speech to Text (STT) solutions for specific customers and use cases were not possible.

Scott clarifies what he means by accuracy: “There are little things like making sure your company name is transcribed correctly. To bigger things where if you change one word in a sentence or you miss one word, it can change the entire interpretation of a sentence.” Good data is not just defined by the overall accuracy or word error rate (WER) but by the accuracy of important, content-heavy words in the conversation. For their analysis, getting all the “a”, “the”, and “of” words right doesn’t matter as much as getting the important words: product or company names like “XL40” and “” or sentiment words like “great” or “poor”.

The Solution

Tailored Speech Models That Perform Better than Big Tech

Before selecting Deepgram, compared many ASR providers including the Big Tech STT providers. Their models were initially a bit better, but with Deepgram, they were able to train and customize a speech model using audio from’s own platform, which led to a dramatic accuracy increase beyond what the all-in-one ASRs could reach. This training was not a resource or time burden for They trained tailored speech models for their customers within weeks with no resources needed from except providing 20+ hours of audio.

The quality of your transcript determines the quality of the information you can extract from its text. Having a customized speech model literally pays dividends on all natural language processing that happens downstream.”

Scott Hoch

Head of Data,

The Results

Accurate Insights for Sales and Sales Management

The insights that’s platform provides are huge value drivers for their customers, but that all comes back to highly accurate transcripts ASR for their machine learning platform. Scott notes, “Your topic modeling, your sentiment analysis, your detection of product features and competitors, and value drivers all come from the transcript. The quality of your transcript therefore has implications all the way down your value chain.”

The end goal of’s customers is to accurately understand, from all the sales data and sales conversations, what drives success and what drives failure. found that creating tailored speech models provides better data to answer these questions for the customer. Their trained models provided their customers with an 18% relative reduction in WER and a general accuracy close to 88%.

Looking Ahead

Expand Overseas to Multilingual Sales Organizations

As expands to multilingual overseas sales reps, they want to reduce any friction to recording and transcribing sales calls. They want to work with Deepgram on automatic language detection so that the rep just picks up the phone and the transcription will be done in the language they or their customer is speaking. This is especially important in Europe as they are always talking with each other but often in different languages.

Overall, Scott says, “It’s been a great partnership.”

View Podsights Case Study