Impact
More accurate answers: cleaner audio gives the model more accurate inputs.
Voice that works in noisy places: the model focuses on the person speaking, even in a busy coffee shop.
Fast integration: a prototype running within a week.
Company background
Perplexity is an AI agent platform focused on accuracy. Its mission is to power the world's curiosity, answering questions and completing tasks by text or voice in its mobile and web apps.
Brady Murphy is Perplexity’s Multimodal Lead. He describes his job as giving people a natural way to interact with AI systems. Before Perplexity, he built video players at YouTube and Prime Video that streamed to TVs and mobile devices around the world.
The challenge
"Delivering accurate answers is our number one priority," Brady said. "And with voice, when you don't have the proper data going into the model, you're going to degrade the accurate answers that are coming back out."
People use voice where they are, on noisy streets and in offices. Whatever is said around them goes into the model too. Brady’s example: “If I ask a question of, what is the weather today in San Francisco? And somebody says New York behind me and it searches, what is the weather in New York? It’s really frustrating.”
A loud voice nearby could also cut off a long question halfway, so the agent never answered it. And one bad first try was often enough to lose a user: "When someone uses voice for the first time and receives an inaccurate answer, you tend to see them stop using it over a period of time, because they feel like it's just not as capable as text input is."
Before ai-coustics, Perplexity used the standard audio processing module common in live streaming, with noise suppression and acoustic echo cancellation. It removed the agent’s own voice from the microphone, but it couldn’t reduce speech and volume from other people nearby.
The team made reducing dissatisfaction in voice interactions a priority. "We really needed to find a solution that could make sure what was going into the model was what the user wanted the model to actually hear."
The integration: from a first conversation to production
Perplexity met the ai-coustics team at a community event in San Francisco. Once the teams started working together, a prototype was integrated within a week. “It really showed that the ai-coustics product could deliver the solution that they promised.”
The speed comes from how the ai-coustics SDK is built. It runs inside Perplexity’s multimodal SDK, and Brady describes the integration as very simple: “Download the model, collect the weights inside of the processor, and then just run your samples through it and it produces samples back out.”
Brady spent years building SDKs at YouTube and Prime Video that many different clients used. His verdict: “The [ai-coustics] SDK was amazing to work with, it was really easy to integrate a solution that already fit well inside of a Rust codebase.”
What changed with ai-coustics
Quail Voice Focus now ships in the multimodal SDK that is delivered to all users’ phones and various desktop applications. It processes audio for speech recognition and voice mode. Whenever someone talks to the app on their phone, ai-coustics is processing the audio.
“When a lot of people are talking around you, it’s going to focus on you and just hear you. It allows you to speak naturally within our product and not just be so confined to a text box."
Brady Murphy, Multimodal Lead, Perplexity
More accurate answers. “What changed for our product was users got more accurate inputs into their interaction with AI models,” Brady said.
Better dictation. "Whenever you're doing dictation or speech recognition, the words that are getting transcribed are going to be more accurate."
Voice in noisy places. "Users that are using voice, they get a better experience when they're within noisy environments."
Accessibility. “Users that maybe need greater accessibility, they don’t have to have a degraded product experience just because they aren’t using a keyboard.”
Why it matters
Audio is one of the most natural ways to interact with technology. "It's really hard to look down at your phone when you're walking around," Brady said. “Being able to interact with an AI using audio is extremely important. That’s how we naturally talk with each other.”
As more people talk to their devices instead of typing on them, AI has to work wherever they are. Teaching machines to hear is what ai-coustics does as we build the audio intelligence layer that production AI runs on.
See how Quail Voice Focus can improve your voice agent.
