Sold by: PolyAI
Deployed on AWS
Dialog-RSN-1 is PolyAI's frontier audio-native model designed for mission-critical enterprise conversations. It reasons directly over raw caller audio and replies through a separate, controllable voice layer, responding in under 300ms median. It achieves the highest scores of any real-time-capable model on quality, safety, and audio understanding on PolyAI's Dialog-Eval benchmark.
Overview
Dialog-RSN-1 is PolyAI's most advanced frontier audio-native model, purpose-built for mission-critical enterprise conversations. It reasons directly over raw caller audio, not a transcript, and replies through a separate, controllable voice layer. It fuses turn-taking, speech recognition, and response into one model, and scores highest of any real-time-capable model on quality, safety, and audio understanding on PolyAI's Dialog-Eval benchmark. In production, that's already showing up as results: a national restaurant group saw an 11% relative increase in containment after switching to Dialog-RSN-1.
Every other voice architecture makes a trade-off. Cascaded pipelines (ASR to LLM to TTS) throw the audio away: the model only ever sees a best-guess transcript, so tone, hesitation, and uncertainty are gone before it can reason. Speech-to-speech models keep the audio but give up control: the output voice is baked in, with no way to fix mispronunciation. Dialog-RSN-1 gives up neither: audio-native reasoning on the input side, with speech generation left to a separate, controllable voice layer.
Supports English voice calls today, with additional languages on the roadmap. Deployed as a SageMaker model directly in your own AWS account, with no third-party inference dependency. ISO/IEC 27001:2022 and Cyber Essentials Plus certified; SOC 2+ Type 2 audited; HIPAA Security Rule controls in place (BAA available) and PCI DSS controls in place; GDPR-ready with EU residency.
Highlights
- Fastest reasoning model built for live calls. The model's own first-token response is under 300ms median. Across the full caller exchange, production data puts it at about 390ms typical vs. 700-730ms for other real-time voice architectures. At one large insurance provider, switching to Dialog-RSN-1 cut response latency by 37%.
- Hears what a transcript can't. Reasons over raw audio, not a transcript, catching hesitation and knowing exactly when to speak. It keeps the full conversation in view: fixing misheard names, telling a human from an answering machine, with no per-flow ASR tuning or vocab files.
- Enterprise control, zero lock-in. Independent guardrails run at every layer: input screening, grounded answers, output filters, and full audit logs. The same agent, tools, and knowledge run on Dialog-RSN-1, Raven 3.5, or a cascaded pipeline, so you can A/B test or fall back without rebuilding.
Details
Sold by
Categories
Delivery method
Latest version
Deployed on AWS
New
Introducing multi-product solutions
You can now purchase comprehensive solutions tailored to use cases and industries.
Features and programs
Financing for AWS Marketplace purchases
AWS Marketplace now accepts line of credit payments through the PNC Vendor Finance program. This program is available to select AWS customers in the US, excluding NV, NC, ND, TN, & VT.
Pricing
Pricing is based on actual usage, with charges varying according to how much you consume. Subscriptions have no end date and may be canceled any time.
Additional AWS infrastructure costs may apply. Use the AWS Pricing Calculator to estimate your infrastructure costs.
Dimension | Description | Cost/host/hour |
|---|---|---|
ml.g6e.4xlarge Inference (Batch) Recommended | Model inference on the ml.g6e.4xlarge instance type, batch mode | $120.00 |
ml.p4d.24xlarge Inference (Real-Time) Recommended | Model inference on the ml.p4d.24xlarge instance type, real-time mode | $45.00 |
ml.p5.48xlarge Inference (Real-Time) | Model inference on the ml.p5.48xlarge instance type, real-time mode | $110.00 |
Dimensions summary
You pay by the host hour, based on which compute instance runs the model and how it processes calls. One option uses the ml.g6e.4xlarge instance in batch mode, which handles requests in groups rather than live. The other two run in real-time mode for live calls, using either the ml.p4d.24xlarge or the ml.p5.48xlarge instance. Your cost scales with how many hours each instance stays active. Choose the instance and mode that match your workload; larger instances and real-time processing bill at their own hourly rates.
Top-of-mind questions for buyers
Real-time mode runs the model on demand for live calls, where turn-taking and latency matter. Batch mode processes requests in groups rather than live. Both bill per host hour. You pick the mode that fits your workload, then pay for each hour the chosen instance stays active.
A host hour is one hour of runtime for the compute instance running the model. Each of the three options maps to a specific instance type: ml.g6e.4xlarge, ml.p4d.24xlarge, or ml.p5.48xlarge. You pay for every hour that instance stays active, regardless of how many calls it processes in that hour.
Each dimension bills independently per host hour. Running multiple instances means each accrues its own hourly charge in parallel. Your total cost is the sum of active hours across all running instances. Choose a single instance for a workload, or combine instances to handle different call types or modes.
poly.ai
Helpful?
Vendor refund policy
Model charges are billed based on actual usage (per hour of inference). We do not offer refunds for usage charges except in cases of billing errors or defects that prevented the product from functioning. To request a refund, contact developers@poly-ai.com with your AWS account ID, product ID and invoice details.
Legal
Vendor terms and conditions
Upon subscribing to this product, you must acknowledge and agree to the terms and conditions outlined in the vendor's End User License Agreement (EULA) .
Content disclaimer
Vendors are responsible for their product descriptions and other product content. AWS does not warrant that vendors' product descriptions or other product content are accurate, complete, reliable, current, or error-free.
Delivery details
Amazon SageMaker model
An Amazon SageMaker model package is a pre-trained machine learning model ready to use without additional training. Use the model package to create a model on Amazon SageMaker for real-time inference or batch processing. Amazon SageMaker is a fully managed platform for building, training, and deploying machine learning models at scale.
Deploy the model on Amazon SageMaker AI using the following options:
Version release notes
Dialog-RSN-1, version 1.2. New model weights (the checkpoint PolyAI serves in production as of 8 October 2026), trained with more turn-taking data. Turn detection scores every frame with both of Silero's 16 kHz and 8 kHz modes and takes the higher, so telephony callers whose voices the 16 kHz model misses are heard at once, including phone audio that Pipecat or LiveKit upsample to 24 kHz before sending. Protocol ids stay within 32 characters. Responses carry status_details and usage.output_token_details, as Pipecat expects. The client's User-Agent is logged on connect. Same instance types and pricing as 1.1.
Additional details
Inputs
- Summary
Input is a stream of OpenAI Realtime client events over SageMaker's bidirectional-streaming invocation (InvokeEndpointWithBidirectionalStream). The session is configured once with session.update: system instructions, an optional list of function tools with JSON-schema arguments, optional knowledge-base topics via the poly_knowledge extension, and the audio input format. Caller audio is then sent as base64 PCM16 mono chunks in input_audio_buffer.append. Results of tool calls are returned to the model with conversation.item.create (function_call_output). The model reasons over the audio itself, so tone, hesitation and non-speech cues reach the model without a separate transcription step. Text-only input via conversation.item.create (input_text) is also accepted.
- Limitations for input type
- Audio: PCM16, mono, 16 kHz or 24 kHz only (audio/pcm); G.711 mu-law/A-law, Opus and stereo are not accepted. Context window: 32,000 tokens across instructions, tools, knowledge topics, history and audio; there is no automatic truncation, so long sessions should be managed by the client. Session: maximum 30 minutes per connection, with an idle timeout; state is per connection and is not carried across reconnects. Language: English. Output modality is text only; a request for audio output is rejected.
- Input MIME type
- application/json
{"type":"session.update","session":{"type":"realtime","model":"dialog-rsn-1","output_modalities":["text"],
"instructions":"You are Sam, a voice assistant for Maison Lumiere, a restaurant. If the caller asks to book, call check_availability before answering. Keep answers to one short sentence.",
"audio":{"input":{"format":{"type":"audio/pcm","rate":16000},"turn_detection":{"type":"server_vad"}}},
"tools":[{"type":"function","name":"check_availability","description":"Check whether a table is free at a given time.",
"parameters":{"type":"object","properties":{"party_size":{"type":"integer"},"date":{"type":"string"},"time":{"type":"string","description":"24-hour HH:MM"}},"required":["party_size","time"]}}]}}
{"type":"input_audio_buffer.append","audio":"<base64 PCM16 chunk>"}
{"type":"input_audio_buffer.append","audio":"<base64 PCM16 chunk>"}
{"type":"conversation.item.create","item":{"type":"function_call_output","call_id":"call_d3200c12","output":"{\"available\":true,\"party_size\":4,\"time\":\"19:30\",\"table\":\"window\"}"}}
Input data descriptions
The following table describes supported input data fields for real-time inference and batch transform.
Field name | Description | Constraints | Required |
|---|---|---|---|
session.instructions | string
System prompt: role, scope, business rules, turn-taking guidance | No | |
session.tools[ ] | function[ ]
JSON-schema function definitions the model may call | - | No |
session.poly_knowledge[ ] | object[] Knowledge-base topics {name, content, actions} the model answers from | - | No |
session.audio.input.format | object {"type":"audio/pcm","rate":16000|24000} | - | No |
session.audio.input.turn_detection | object or null server_vad (default) or null for push-to-talk via input_audio_buffer.commit | - | No |
session.output_modalities | ["text"] Fixed; audio output is refused | - | No |
session.max_output_tokens | integer Optional cap on reply length | - | No |
input_audio_buffer.append.audio | base64 string PCM16 mono chunk | - | No |
conversation.item.create.item | object function_call_output or input_text | - | No |
Resources
Support
Vendor support
For technical questions, integration assistance, or issue reports, contact the PolyAI developer support team at developers@poly-ai.com . Join the PolyAI Discord community at https://discord.gg/xKS3EcS7H to connect with other developers, ask questions, and share feedback on Dialog-RSN-1.
AWS infrastructure support
AWS Support is a one-on-one, fast-response support channel that is staffed 24x7x365 with experienced and technical support engineers. The service helps customers of all sizes and technical abilities to successfully utilize the products and features provided by Amazon Web Services.
Similar products
Omnichannel conversational AI agents that resolve customer enquiries across voice and chat, understanding every caller regardless of accent, dialect, or language. Built on Raven, PolyAI's proprietary customer-service LLM, and deployed by enterprises handling up to 90% of customer calls without a human agent.
The Dialog API uses symbolic techniques to form a control plane, accelerator, and LLM router for text, voice, and multimodal inputs.
Turn every meeting into a searchable knowledge asset. Enterprise-grade security, flexible data-residency options, and deployment delivered by AudioCodes.
AI-powered platform with chat support, enabling users to ask data questions, create visualizations, and generate insights for informed decision-making.
Customer reviews
No customer reviews yet
Be the first to review this product . We've partnered with PeerSpot to gather customer feedback. You can share your experience by writing or recording a review, or scheduling a call with a PeerSpot analyst.