Kensho Scribe

S&P Global is a leading provider of financial data, including credit ratings and analytics for global capital and commodity markets.

The company catalogs tens of thousands of corporate events and their corresponding metadata every day, including buybacks, guidance, defaults, board meetings and more than 150 other event types. In total, about 40,000 of these events are derived from audio files each year, essentially anything you could dial into or listen to – earnings calls, management presentations, acquisition calls and more.

S&P Global employs hundreds of rigorously trained analysts who have created over 100,000 hours of transcripts of these events over the past decade. The transcripts are 99 percent accurate given their strict service level agreements. Their clients use these transcripts to quickly find the information they need to confidently make decisions. It’s much more efficient to simply search a transcript of an earnings call than it is to listen to the whole event.

…transcripts are 99 percent accurate given our strict service level agreements. Our clients use these transcripts to quickly find the information they need

The organization wanted to streamline the process to free analysts’ time for more value-added activities, but none of the existing commercial speech recognition solutions were up to the task. Too many of the audio clips included poor quality, financial or business jargon, accented speech or other challenges that resulted in errors in the transcriptions. This is how the idea for Kensho Scribe was born.

With a motto of “Try crazy things and make them real,” the Kensho Applied R&D team embarked on the process of building a better machine-based transcription service.

The team used the trove of audio and corresponding near-perfect transcripts as a training corpus. More than 100,000 hours of high-quality, domain-specific, labeled data. A dream come true for any machine learning research scientist.

Kensho Scribe saves 1.25 hours per call relative to the legacy process, with more than 50,000 person hours saved to date. The unlocked bandwidth has enabled us to increase transcription coverage by more than 1,500 companies…

One year later, Kensho Scribe was being used to transcribe 99 percent of the earnings calls, management presentations and acquisition call transcripts that S&P Global creates each year. Kensho Scribe saves 1.25 hours per call relative to the legacy process, with more than 50,000 person hours saved to date. The unlocked bandwidth has enabled S&P Global to increase transcription coverage by more than 1,500 companies and expand to other high-impact uses like transcribing voicemails, powering compliance efforts, creating meeting minutes and processing interviews.

In 2020, we opened the next chapter in the story by beginning to make Kensho Scribe available to the rest of the world. As we expand to new use cases for the product, we will continue to push the envelope of transcription quality, evolving the machine learning algorithms that make Kensho Scribe so powerful and accurate.