From understanding Indian English to understanding conversations.
PivotX began training speech-recognition models in 2019, before AutoQC. The challenge was practical: make recorded speech useful when strong Indian English accents and specialized vocabulary made transcription difficult.
2019: training for the speech we actually heard
We trained QuartzNet models to handle strong Indian English accents, then deployed them across multiple speech-recognition use cases. Our focus was on the recordings people needed to process, including speech that general-purpose systems struggled to recognize.
One deployment: tens of thousands of hours of IIT lectures
For NPTEL lecture transcription, our models processed tens of thousands of hours of IIT lectures. These recordings combined strong accents with technical vocabulary and uncommon terms—the long tail of words that matters when someone is learning a specialized subject.
In our 2020 evaluations on this lecture content, our models delivered over 93% transcription accuracy. Leading commercial models we evaluated on the same content delivered roughly 60–70% accuracy: about one word wrong in every three, and sometimes more.
This was a result on that lecture content and evaluation setup. The transcripts supported downstream translation workflows for NPTEL.
Speech recognition became speech and intelligence
Getting the words right was a foundation. The next step was understanding what those words meant for the people using the recordings.
We worked with market research agencies on research use cases: processing interviews and responses, checking fieldwork, and extracting findings relevant to a study. We also worked on contact-centre use cases for multiple brands, bringing recorded conversations into quality review and business analysis.
Across these workflows, the questions extended beyond transcription: which concern was raised, whether a required step happened, what customers were asking for, and which findings needed a closer review.
That experience continues in AutoQC
AutoQC is a product of PivotX. It brings that focus to recorded calls, field conversations and research responses, with analysis configured around the workflow, its context and the result the team needs.
A transcript makes speech readable. AutoQC helps teams examine the conversation against their own process and questions, including how a customer relationship develops across several conversations.
Our earlier speech-recognition work is part of that history. Current model and pipeline choices depend on the requirements of each workflow.
What would you like to learn from your recordings?
Explore the use cases or describe your workflow.
Explore use cases