Home / Blog / PolyAI Dialog-RSN-1 is now on Amazon SageMaker JumpStart
Tech News

PolyAI Dialog-RSN-1 is now on Amazon SageMaker JumpStart

PolyAI released Dialog-RSN-1 on Amazon SageMaker JumpStart, enabling teams to deploy real-time voice reasoning directly within their private AWS accounts.

By Dillip Chowdary • Nov 20, 2026 • Source: AWS Machine Learning Blog

PolyAI Dialog-RSN-1 is now on Amazon SageMaker JumpStart
Image: AWS Machine Learning Blog

At a glance

Product
PolyAI Dialog-RSN-1
Platform
Amazon SageMaker JumpStart
Audio format
16-bit PCM mono, 16 or 24 kHz
Supported languages
English only
Session limit
Up to 30 minutes
Network protocol
OpenAI Realtime protocol

PolyAI released Dialog-RSN-1 on AWS Marketplace and Amazon SageMaker JumpStart, bringing its specialized voice reasoning model into managed enterprise infrastructure according to AWS Machine Learning Blog's report. The deployment allows engineering teams to subscribe through existing commercial agreements and spin up managed real-time endpoints directly inside their private cloud environments. Rather than piping telephony streams through multiple disjointed transcription and natural language systems, the software processes incoming conversational voice in a single unified execution cycle.

This release covers the complete deployment architecture, network protocol translation requirements, latency tuning profiles, and security controls needed to operate enterprise voice agents. Telephony architects, contact center systems engineers, and applied artificial intelligence teams can use the hosted endpoint to maintain private conversational interfaces while retaining complete architectural authority over routing mechanisms and back-office integration.

01Audio pipelines and telephony wire protocols

Engineering teams retain complete architectural control over underlying telephony connections, call routing, and upstream business logic when deploying the system. Dialog-RSN-1 replaces the conversational reasoning tier by processing live voice streams without requiring handoffs across intermediate transcription software layers. Customers connect existing real-time clients directly through the integrated network bridge by reconfiguring their destination base web address to hit the provisioned managed cluster.

Incoming telephony streams require specific preprocessing before reaching the active inference containers. The model ingests linear uncompressed audio streams formatted specifically as sixteen-bit single-channel pulse-code modulation sampled at either sixteen or twenty-four kilohertz. Production telecommunications gateways running standard voice network codecs such as G.711 or Opus must transcode media packets prior to transmission.

Note: Active connection sessions carry a hard technical ceiling of thirty minutes per call, and conversational processing remains restricted strictly to English speech inputs during this initial release.

02Console workflows and endpoint deployment steps

Engineering groups can discover, test, and deploy conversational instances straight from the managed administrative console without building custom operating system containers. SageMaker JumpStart packages baseline serving configurations, driver dependencies, and sample notebooks together so developers can examine conversational behavior on recorded enterprise calls before committing production capital. Provisioning runs through standard interactive console steps or automated Python build pipelines.

  1. Catalog discovery Open the Model Hub inside the Amazon SageMaker AI console and locate Dialog-RSN-1 within the JumpStart catalog.
  2. Endpoint launch Select the target model card and choose the deploy action to stand up an optimized real-time serving cluster.
  3. Pipeline automation Provision programmatic deployment definitions using sample automation scripts provided in the official SageMaker Python SDK repository.

Once activated, administrative teams manage the endpoint alongside standard machine learning workloads within their existing infrastructure dashboard. Automated deployment scripts enable continuous integration systems to stand up isolated test endpoints, benchmark conversational outputs against recorded customer calls, and destroy compute nodes after test suites complete execution.

03Direct audio reasoning and conversational speed

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Traditional voice architectures introduce noticeable latency by routing speech through separate transcription models before feeding intermediate text into language engines for conversational timing decisions. Dialog-RSN-1 processes the caller's acoustic stream directly in a single calculation step, deriving semantic meaning, caller sentiment, and vocal cadence simultaneously. This unified architecture enables conversational systems to respond immediately without suffering from accumulated pipeline processing delays.

Direct acoustic evaluation allows the model to detect conversational hesitation, emotional frustration, and natural speaking pauses before conversational tension escalates. By analyzing speech patterns directly from audio waveforms rather than relying on flat textual transcripts, the agent determines the appropriate conversational moment to speak. Callers experience natural dialogue pacing that mirrors interactions with experienced contact center staff.

Acoustic processing preserves critical conversational capabilities during periods of high telecommunications volume. The system maintains continuous barge-in responsiveness even when call volume spikes, preventing the performance degradations that force traditional pipelines to disable conversational interruption features. In addition, the engine generates output transcripts directly from the source audio rather than relying on external approximations.

04Preconfigured serving profiles and performance metrics

System administrators can select purpose-built optimization profiles inside the deployment console instead of manually benchmarking dozens of compute instances. These target configurations balance infrastructure costs against processing latency based on specific operational requirements.

Optimization profileCore operational objectivePrimary production use case
Latency optimizedMinimizes Time to First TokenInteractive real-time voice calls
Throughput optimizedMaximizes tokens processed per secondHigh-volume automated contact centers
Cost optimizedMinimizes compute expenditure per requestAsynchronous or non-urgent audio processing
BalancedEqual tradeoff between speed and costGeneral-purpose enterprise voice workloads

Before provisioning compute resources, the console interface displays concrete operational benchmarks including median latency distributions, Time to First Token metrics, and aggregate token throughput. Telephony architects use these surfaced indicators to select the right compute cluster footprint, eliminating time-consuming manual load tests.

05Account boundary isolation and compliance governance

Inference workloads run completely inside the customer's private virtual private cloud rather than routing voice data through multi-tenant external endpoints. Customer audio payloads, dynamic model outputs, and generated text transcripts never leave the organization's verified cloud environment. This deployment architecture provides complete data residency enforcement and auditability for organizations subject to stringent regulatory oversight.

Access control relies entirely on existing identity policies, role permissions, and managed security boundaries already configured within the host enterprise environment. Operational engineering teams monitor endpoint health, invocation volume, and inference latency using native CloudWatch metric collectors and centralized system event logs.

Billing consolidates directly through existing enterprise agreements and regular procurement channels on the AWS Marketplace. Organizations deploy advanced voice reasoning systems using the identical access controls, monitoring dashboards, and compliance postures already governing their broader cloud application portfolio.

Frequently asked questions

What audio formats does Dialog-RSN-1 accept?

The endpoint accepts 16-bit PCM mono audio sampled at either 16 kHz or 24 kHz. Telephony pipelines streaming G.711 or Opus codecs must transcode caller audio before transmitting it to the model.

Where does inference compute take place?

Inference runs entirely within the customer's private AWS account, inside their own VPC and behind their existing IAM controls. Call audio and generated transcripts never leave the customer's managed cloud boundary.

How does the model connect to existing voice architectures?

The endpoint uses the OpenAI Realtime protocol, allowing existing Realtime clients to route traffic through PolyAI's bridge by updating the destination base URL. Telephony orchestration and business logic remain under customer management.

What optimization profiles are available for deployment?

JumpStart provides four distinct optimization profiles: Latency optimized for interactive calls, Throughput optimized for high call volumes, Cost optimized for minimal compute expense, and Balanced for general-purpose workloads.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →