PolyAI’s Dialog-RSN-1 can now be deployed as a SageMaker endpoint through AWS Marketplace or SageMaker JumpStart. The model runs inside the customer’s AWS account, under its virtual network, identity policies and monitoring, so call audio and transcripts remain within an environment the organization controls.

The model processes live audio directly rather than chaining separate speech recognition and language models. PolyAI says this lets it use timing, hesitation and tone when deciding what a caller means and when to respond. Existing clients built for the OpenAI Realtime protocol can connect through a bridge by changing their base URL.

Deployers still own the surrounding system: telephony, business logic, orchestration and escalation. Input must be 16-bit mono PCM at 16 or 24 kHz, other formats require transcoding, connections are limited to 30 minutes, and the model currently works only in English.

JumpStart provides preconfigured latency-, throughput-, cost- and balanced deployment profiles, with measured latency and throughput shown before launch. The simplified setup removes container and instance selection, but buyers should evaluate the model on their own call mix, accents and interruption patterns before placing it in a customer-service workflow.