Enterprise leaders deploying voice agents for customer service now have a third architectural option, and it closes a gap that cascaded pipelines and full speech-to-speech models have each struggled to fill.
On July 30, 2026, PolyAI published a technical blog post introducing Dialog-RSN-1: an audio-native large language model that fuses turn-taking, speech recognition (ASR), function calling, and response generation into a single model. Unlike cascaded systems (where audio flows through a chain of separate models) or full speech-to-speech models (which generate audio output too), Dialog-RSN-1 keeps the output voice in a separate text-to-speech system. The result: a model that hears calls the way a skilled human agent would, while still letting enterprises tune exactly what callers hear.
The model is already handling live calls in production for PolyAI’s existing enterprise customers. CMSWire, SiliconANGLE, and TechTimes all covered the launch. Here is what enterprise AI and GTM leaders need to understand.
What Dialog-RSN-1 Actually Does
Most enterprise voice agents today work through a cascade: a voice activity detector decides when the caller has stopped speaking, an ASR model converts audio to text, an LLM processes that text and calls tools, and a TTS system speaks the response. Every hop adds latency and discards signal. By the time the LLM reasons over the transcript, tone, hesitation, and emotional cadence are gone.
Full speech-to-speech models (like OpenAI GPT Realtime) solve part of this by processing audio end-to-end, but they generate audio output directly, which constrains how enterprises control the voice callers hear. Swapping voices requires model retraining, not a configuration change.
Dialog-RSN-1 takes a third path. Per PolyAI’s technical writeup:
- The model ingests raw audio on the input side, reasoning over the full acoustic signal without requiring a prior ASR pass.
- Turn-taking is the model’s very first output token:
EMPTY(no speech detected),ONGOING(caller is still speaking), orCOMPLETE(ready for a reply). This means the model commits to whether to speak before generating a word of response. - Function calling, tool use, and response generation all happen within the same model that heard the audio.
- Speech output is handed to a separate, externally configurable TTS system, so enterprises retain full control over voice, accent, and emotional tone without touching the core dialog model.
PolyAI built Dialog-RSN-1 through supervised and reinforcement fine-tuning on top of open-weight multimodal base models, evaluating architectures including Gemma, GPT-OSS, Qwen, and Mistral. The target inference profile is 8B dense to 30B sparse parameters on A100 GPUs to achieve the sub-300ms latency envelope.
Architecture Comparison for Enterprise Voice AI
The three dominant enterprise voice AI architectures each carry distinct trade-offs:
| Architecture | Input Side | Turn-Taking | Speech Recognition | Output Voice |
|---|---|---|---|---|
| Cascaded pipeline | Text transcript | Separate VAD | Separate ASR model | Separate TTS (fully controllable) |
| Full speech-to-speech (e.g. GPT Realtime) | Raw audio | Integrated | Integrated | Integrated (limited control) |
| Dialog-RSN-1 | Raw audio | Integrated | Integrated | Separate TTS (fully controllable) |
Cascaded systems are easiest to swap out and audit, but they sacrifice audio-side signal at every hand-off. Full speech-to-speech models reclaim that signal but lock enterprises into the vendor’s output voice. Dialog-RSN-1 is designed to provide audio comprehension on the input side with output voice independence.
Production Latency Numbers for Enterprise Voice AI
PolyAI reports the following from live deployments, corroborated by MarkTechPost’s technical summary:
- Sub-300ms end-to-end latency for Dialog-RSN-1 in production, with a tighter distribution than prior systems.
- 860-1,900ms for GPT Realtime 2.1 in the same production environment, according to PolyAI’s live deployment data.
- +11% relative containment improvement at a restaurant group customer, meaning more calls resolved without escalation to a human agent.
- -37% latency reduction at an insurance customer compared to the prior cascaded architecture.
Human conversational turn-taking typically involves pauses of 200 to 300ms. Response latencies above 500ms are perceivable; above 1,000ms they degrade caller experience in measurable ways. The gap between 280ms and 1,000ms is the difference between a conversation that feels fluid and one that feels like an automated phone tree.
On Dialog-Eval, PolyAI’s internal benchmark (which the company plans to open-source), Dialog-RSN-1 outperforms the cascaded-model ceiling of 77 and scores above GPT Realtime 2.1 on audio-aware examples. The benchmark evaluates each example at a single decision point in a call, scoring the atomic next step rather than rolling out a full conversation.
What Enterprise Buyers Need to Weigh
Access model. Dialog-RSN-1 is not available as a public API or open weights. Existing PolyAI customers can enable it in their projects today. New customers must request early access. PolyAI operates on custom enterprise contracts; third-party estimates (cited by TechTimes) place these in the range of $100,000+ annually plus per-minute usage fees. This positions Dialog-RSN-1 in the same tier as NICE CXone and Cognigy, not as a developer commodity.
Language coverage. English only at launch. The underlying model is multilingual, and PolyAI reports it is already detecting non-English speech in production. Multilingual rollout is described as the next priority. For non-English contact center use or rich web chat, PolyAI’s Raven 3.5 remains the recommendation.
Deployment context. PolyAI now serves 100+ enterprise customers with 2,000+ live deployments across restaurants, insurance, financial services, healthcare, hotels, retail, telecom, and travel. The company closed an $86M Series D in December 2025. That base makes Dialog-RSN-1 a production-grade offering from a company with real scale, not a research preview.
Competitive dynamics. The enterprise voice AI market is maturing from a single-vendor question to an architectural one. Enterprises that signed up for cascaded systems in 2024 are now evaluating whether audio-native comprehension is worth a platform migration. Those that went full speech-to-speech are evaluating whether output voice control matters enough to justify the latency and flexibility trade-offs. Dialog-RSN-1 is designed as the answer to both questions simultaneously.
This is not a standalone development. The broader enterprise voice AI race between OpenAI and Anthropic in July 2026 set the competitive backdrop: both labs are investing heavily in voice-capable agents for knowledge work. PolyAI’s move positions specialist enterprise voice infrastructure against both general-purpose AI stacks. Enterprise leaders building agentic AI execution layers for revenue and operations will need to decide whether contact center voice is best served by a horizontal AI platform or a specialized voice-native stack.
The Practical Signal for Enterprise AI Leaders
The architecture choice for enterprise voice agents is no longer just about which LLM to use. Dialog-RSN-1 makes the case that the audio-comprehension layer itself is a meaningful capability surface, separate from both the reasoning model and the output voice. Latency below 300ms in production, coupled with improved containment metrics and retained TTS control, is the kind of operational argument that resonates with CX leaders and AI infrastructure teams who have been burned by cascaded pipelines losing emotional context.
For enterprises evaluating voice AI investments in Q3 2026, the decision tree has a new branch. Existing PolyAI customers should evaluate Dialog-RSN-1 for their highest-volume English-language contact flows. Enterprises on cascaded architectures with latency or containment problems have a validated alternative to full speech-to-speech. And procurement teams should note that this segment is now consolidating around enterprise-grade contracts with meaningful per-deployment commitments.
If your organization is assessing where enterprise voice AI fits in your broader AI transformation, start with a conversation.