The Pulse
AWS customers can now deploy PolyAI's voice-dialog model through SageMaker JumpStart
AWS customers can deploy PolyAI's voice-dialog model as a SageMaker endpoint through JumpStart. The model accepts live-call audio, while customers retain responsibility for telephony integration and business logic.

AWS customers can now deploy PolyAI's Dialog-RSN-1 voice model from the Amazon SageMaker JumpStart catalog as an endpoint in their own accounts. The model handles live-call audio and generates responses for customer-conversation agents, according to an AWS post published Nov. 19.
The JumpStart listing follows an earlier SageMaker deployment route: PolyAI said on Oct. 6 that customers could subscribe through AWS Marketplace and deploy the model as a SageMaker endpoint. The new catalog entry provides deployment and evaluation options within SageMaker AI.
Deployment and integration
Customers can subscribe through AWS Marketplace or deploy from the JumpStart catalog in SageMaker. In the SageMaker AI console, they select the model from the hub and choose Deploy; the resulting real-time inference endpoint uses a preconfigured serving stack.
The endpoint runs within the customer's AWS account and virtual private cloud, with access through AWS Identity and Access Management roles and monitoring through Amazon CloudWatch. Customers can also use the SageMaker Python SDK for automated deployment workflows.
The endpoint uses the OpenAI Realtime protocol, and an existing Realtime client can connect through PolyAI's bridge by changing its base URL. Input audio must be 16-bit PCM mono at 16 or 24 kHz; G.711 or Opus audio must be transcoded first. Connections can last up to 30 minutes, and the model currently supports English only.
The customer retains responsibility for orchestration, telephony integration and business logic. The endpoint is billed through the customer's AWS account.
Testing and serving choices
The catalog includes an overview and sample notebooks. After deploying an endpoint, a customer can use PolyAI's sample notebook to stream recorded caller audio and print the model's turn decisions, replies and transcripts.
JumpStart offers a set of optimization profiles for the serving stack. The latency-optimized profile targets time to first token for interactive calls, while the throughput-optimized profile targets tokens per second for higher call volumes. Cost-optimized and balanced profiles are also available.
Before deployment, JumpStart displays metrics including P50 latency, time to first token and throughput. The SDK route can also be used to incorporate deployments into continuous integration and continuous delivery workflows.
How the model handles calls
Dialog-RSN-1 takes caller audio directly and combines turn-taking, speech recognition, function calling and response generation in a single model. PolyAI says it can use information such as tone and timing when deciding how and when to respond.
The model does not use a separate speech-to-text handoff before reasoning over the caller's audio. PolyAI says the model is audio-aware on the input side and leaves speech generation to a separate text-to-speech system.
PolyAI introduced Dialog-RSN-1 on July 30 and says it is already handling live calls in production. The company says the model was built through further training of open-weight multimodal models using in-house datasets, including conversations based on real deployments and synthetic conversations.
Sources
- Original source · Blog post Amazon Web Services aws.amazon.com
Other pages read
- PolyAI · Dialog rsn aws sagemaker poly.ai
- PolyAI · PolyAI dialog rsn poly.ai
Companies in this story
Related stories
- OpenTSLM time-series model becomes available in SageMaker JumpStart
- Salesforce uses SageMaker routing to shift model requests across regions
- DeepL launches Aure voice translation model
- SageMaker adds cache-aware routing for real-time AI endpoints
- NVIDIA’s 120B Nemotron 3.5 Super VL is available on SageMaker