Webinar

Speech-to-Text built for telecom

Real-world phone calls are unpredictable. Background noise, weak cell signals, varied accents, and mid-sentence signal issues trip up speech-to-text models built for clean audio. At Reverb®, Bandwidth showed a first look at its in-house speech-to-text model, architected and trained from scratch to run on Bandwidth GPUs in the call path. In this session, Mike Nelson and Jason Sommerset of Bandwidth Labs go deeper on how it was built, how we measure it, and what it means for real-time voice AI.

What we’ll cover:

How we built the model and the inference stack, kernels and all, to run on our own edge GPUs inside the network, with no extra hop and no one else’s cloud.
How we measure it on word error rate and semantic word error rate across three benchmarks: our own set of hard phone audio, the Open ASR Leaderboard, and Pipecat AI’s semantic benchmark. We’ll show the full charts so you can draw your own conclusions.
What streaming-native means in practice: no word revisions, latency holds steady at scale, and you can act on a word while the sentence is still happening, like redacting a card number as it’s read.

Speakers

Mike Nelson

Principal Architect, Innovation

Mike is a Principal Architect on Bandwidth’s Labs team, where he does research and development. His focus spans AI research, model development, and software engineering, backed by hands-on work with GPU hardware. He brings more than 20 years of experience to the role.

Jason Sommerset

Principal, Product Strategy and Innovation

Jason Sommerset leads product strategy for Bandwidth Labs, focusing on Voice AI, speech-to-text models, and the AI infrastructure that runs them on Bandwidth’s network. Equally at home in the technical details and in go-to-market and P&L strategy, he turns early market bets into products customers buy. In more than 12 years at Bandwidth, he helped launch its first developer messaging and voice API platform and its Enterprise BYOC offering for UCaaS and CCaaS, led the global CPaaS product and development team of 60+, and served as GM and SVP of Global Voice Plans.

Speech-to-Text built for telecom

Speakers

Mike Nelson

Jason Sommerset

Related posts

Reverb

Lumen to Bandwidth Migration Guide

Voice AI that wins requires “humanness” at the network level

Mythbusting: What Lumen’s voice stop-sell really means for your business