Anam cuts unwanted avatar interruptions by about half with Deepgram Flux

The turn-taking problem

Anam describes itself as the builder of the "face layer for agents": photorealistic, emotionally expressive avatars that offer sub-second response times and API-first deployment. According to the company, 70% of users prefer them over voice-only and text-based interfaces.

The quality of that experience depends on whether an avatar can hold a natural, real-time conversation. Every few seconds it has to decide whether a person has finished speaking or is just pausing to think. Customers framed this as a problem with the avatars' "ears": the avatar had to feel like it was listening, not only transcribing audio.

Anam also runs latency-sensitive, high-concurrency workloads. That meant any speech-to-text (STT) model had to be deployable on its own infrastructure, rather than adding a network hop to a third-party cloud.

How Anam tested Flux

Anam ran Flux in A/B tests against the option it knew best, a self-hosted open-source STT model. It scored the candidates on transcription response latency and the number of unwanted interruptions. The whole Anam team then held end-to-end conversations with the agent personas.

Requirements were strict: accurate turn-taking across speaking styles, low end-of-turn latency, resilience to background noise, the ability to self-host on Anam's Kubernetes cluster, and concurrency that scales with growth. According to the case study, Flux handled conversational edge cases the open-source model could not.

Anam CTO Ben Carr said the team also compared Flux with its home-grown cascading pipeline and competitor solutions. "Across varied talking styles and audio conditions, Flux consistently outperformed on latency and unwanted avatar barge-ins," he said.

In production, user audio streams into Flux, which transcribes in real time and signals when the user has actually finished their turn. Anam's language model and voice output build on that timing. Deepgram highlights three differences:

• Speakers can pause mid-thought without the avatar jumping in, which helps slower speakers and role-play scenarios.
• End-of-turn detection is fast enough that the avatar responds without dead air.
• Non-user sounds are ignored, so a barking dog no longer counts as a turn.

Deployment and results

Anam deployed Flux self-hosted on its own Kubernetes cluster with H100 GPUs, the same infrastructure used across its platform. After validation, it rolled Flux turn-taking out to 100% of its turnkey users. Using Deepgram's hosted service took days from first test to production, while the self-hosted setup took around two weeks.

The headline results come from Anam's own A/B testing: about 50% fewer unwanted avatar barge-ins and no latency penalty, with p95 flat across tested alternatives. Complaints about turn-taking accuracy on English sessions running Flux dropped to near zero. Carr said that before Flux, Anam "had to choose between slow latency and unwanted avatar interruptions," and that it no longer sees complaints about either on supported flows.

Some customers who had used their own STT migrated to Anam's turnkey experience. Anam notes that customers who bring their own STT sit outside the pipeline, which limits how much conversation quality it can control.

What it means for Anam and its customers

Anam says companies use its avatars for customer-facing AI agents and for training and simulation. It cites gains of 44% in engagement, 2x faster adoption and 25%+ higher conversion when the avatars are added to the sales cycle. The case study does not tie those figures specifically to Flux.

Looking ahead, Anam expects avatar experiences to become more multilingual, expressive and configurable. As Flux expands language coverage, the company expects to use it more broadly by default. It is also evaluating Deepgram's Flux TTS offering for the next generation of voice quality and real-time avatar presence.