Mark Backman
Conference affiliation: Daily · 2025
Related work and projects
- Pipecat Cloud is Now Generally Available ↗daily.co
- Benchmarking STT for Voice Agents ↗daily.co · Source ↗
- Pipecat ↗github.com · Source ↗
Mark Backman led product at Daily and helped drive the development of Pipecat, an open-source framework for building real-time voice and multimodal agents. He built a voice assistant that could answer questions about World’s Fair speakers and ran into a tradeoff that shaped the application. Looking up speaker records mid-conversation took too long, while giving the model those records up front sped up replies at the cost of sometimes confusing similar entries. Daily’s account of Pipecat’s development · World’s Fair voice AI workshop
On this page
On October 29, 2026, when Deepgram announced it was acquiring Daily and establishing an independent foundation for Pipecat, Mark said he was joining forces with Deepgram and would continue working on Pipecat with the foundation’s partners. Mark’s announcement · Deepgram’s announcement
From video calls to voice agents
As vice president of product at Oblong Industries, Mark encouraged managers of remote teams to use video for conversations that build working relationships. At Daily, his work covered streaming video, real-time communications and the SDKs developers use to embed calls in their own products. In 2023, he compared a laptop webcam with an iPhone in the same backlit room, then showed how to select the phone as a camera in Daily’s pre-call screen and explained a Chrome compatibility issue that could get in the way. Oblong and remote work · Daily author profile · Webcam comparison
Pipecat applies that real-time communications work to a call where one participant is software. Its composable audio pipelines let developers swap out speech recognition, language models and speech synthesis independently, or use a single model that takes in and returns audio. The application still manages the connection, conversation state, tools and recordings. Mark recommends WebRTC for interactive client connections because it handles problems such as packet loss that can degrade live audio. For a connection between an agent server and a telephony provider, he favors WebSockets. Pipecat project · World’s Fair voice AI workshop
Mark and Nina Kuruvilla co-wrote the January 2026 general availability announcement for Pipecat Cloud after a beta involving more than 1,000 teams. Daily’s managed platform took on work such as scaling agents as calls came in and running them across regions. It also offered telephony and observability, while developers kept the option of running their Pipecat applications on their own infrastructure. Pipecat Cloud launch
What happens between turns
- Keep lookups off the critical path when the tradeoff is worth it. Mark’s World’s Fair assistant could start playing a streamed reply once it had enough text to speak. A tool call had to finish generating its structured arguments before the lookup could begin. Putting speaker records directly into context removed that wait and made similar records easier to mix up. For task-oriented calls, Mark recommends giving the model one step at a time and dropping details once they’ve served their purpose, such as a date of birth used only for an initial verification. World’s Fair voice AI workshop
- Measure errors that change meaning. In February 2026, Mark reported on the team’s open-source speech recognition benchmark, which initially compared ten services on 1,000 real-world English speech samples. Semantic Word Error Rate discounts harmless formatting differences and counts errors that change what the speaker meant, such as getting a name or an instruction wrong. The benchmark also measures how long each service takes to finalize a transcript after the speaker stops. The 95th and 99th percentiles reveal occasional delays long enough to break the rhythm of a call even when the median is fast. Benchmark report · Benchmark code and methodology
- Give the caller time to finish a thought. Mark described how the Pipecat team’s Smart Turn model addresses the tradeoff in a fixed silence timeout. A short timeout gets the agent talking sooner and can cut off a caller who pauses to think. Smart Turn classifies the input audio as a complete or incomplete turn, letting Pipecat wait longer when the caller sounds like they have more to say. Smart Turn · World’s Fair voice AI workshop
- Track what the caller actually heard. Mark argued that a speech service’s playback report can arrive too late after an interruption. By then the caller may have spoken again and the agent may be preparing a reply on the assumption that its previous sentence played in full. He pointed to word timestamps as a way for the application to track playback as it happens and carry forward the words the caller actually heard. Pipecat discussion on interruptions
Read the topics behind these talks
1 conference talk
Loading more talks and highlights…
