AI Voice Agent technology is moving fast.

If you've been following the industry, you've probably heard a lot about Speech-to-Speech (S2S) models. They can listen to speech and generate speech directly, without the traditional speech-to-text and text-to-speech steps.

It sounds like the obvious next generation of voice AI.

But there is a less exciting—and arguably more important—question for businesses:

What actually works best in a production call center?

In 2026, the answer is still often ASR → LLM → TTS, also known as a cascaded voice architecture.

It may not be the newest architecture, but it has several major advantages for enterprise contact centers:

• Lower and more predictable AI costs
• Mature and widely available models
• Better control over the conversation
• Easier debugging and monitoring
• Flexible choice of ASR, LLM, and TTS providers
• Strong CRM and business-system integration
• Excellent performance for most real-world call center workflows

Native S2S is improving rapidly and can deliver impressive conversational naturalness and latency. But for many enterprise use cases, the extra sophistication isn't necessarily worth the additional cost or reduced flexibility—at least not yet.

Daily's February 2026 voice-agent benchmark reached a similar conclusion: most production voice agents still use text-mode LLMs with cascaded pipelines, while S2S models are closing the capability gap. The benchmark also found GPT-4.1 at 94.9% on its reported instruction-following benchmark versus 86.7% for GPT Realtime. (Daily)

So let's look at the two architectures from a real call center perspective, not just a technology-demo perspective.

First, What Is a Cascaded Voice Architecture?

The traditional AI Voice Agent architecture is:

Each component has a specific job.

What Is Native Speech-to-Speech?

Native Speech-to-Speech takes a different approach.

Instead of:

the model works directly with audio:

The major advantage is that the model can process speech as speech rather than treating text as the primary intermediate representation.

This can help preserve information such as:

• Tone
• Pauses
• Speaking speed
• Emotion
• Prosody
• Interruptions
• Other characteristics of spoken communication

The result can feel very conversational. And for certain applications, that's a significant advantage.

Cascaded vs. S2S: The Practical Comparison

The key word here is today.

S2S may become the dominant architecture in the future.

But in 2026, the cascaded approach remains extremely practical.

Why Cascaded Architecture Has a Major Cost Advantage

This is one of the biggest reasons the cascaded architecture remains attractive for call centers.

A contact center doesn't make ten voice interactions a day. It might make:

• 10,000 calls/day
• 100,000 calls/day
• Millions of minutes/month

At this scale, AI inference cost matters.

And the ability to choose different models for different jobs becomes extremely valuable.

For example, a company could use:

Instead of paying for one expensive end-to-end model for every second of every conversation.

Real-World Pricing Shows the Difference

Microsoft's published Voice Live pricing provides a useful example. For its Pro tier, Microsoft lists:

• Text input: $5.50 / 1M tokens
• Text output: $22 / 1M tokens
• Audio with Azure AI Speech: $17 / 1M input tokens and $38 / 1M output tokens
• Native audio with GPT-4o Realtime: $44 / 1M input tokens and $88 / 1M output tokens (TECHCOMMUNITY.MICROSOFT.COM)

These aren't a universal cost comparison—actual voice-agent costs depend on tokenization, audio duration, model selection, caching, infrastructure, and other services.

But the pricing structure demonstrates an important point:

Native audio models can carry substantially higher model-level usage costs than a carefully optimized text-based pipeline.

For a small voice assistant, that difference may be insignificant.

For a contact center processing millions of minutes, it can become a major operating expense.

Why Does Cost Matter So Much in a Call Center?

Consider an outbound campaign. Imagine an AI agent makes:

1 million calls per month.

Most conversations may be simple:

"Hello, we'd like to remind you about your appointment tomorrow."

or:

"Would you like to schedule a product demonstration?"

You don't necessarily need the most advanced speech-native model available to handle these conversations.

A well-designed cascaded system can use an efficient model for routine interactions and reserve more expensive reasoning capacity for complicated cases.

This creates a powerful architecture:

This is not just an AI architecture decision.

It's an operating-cost strategy.

Cascaded Architecture Can Already Deliver Real-Time Conversations

One common criticism is:

"ASR → LLM → TTS has too many steps. It must be slow."

That was a bigger problem with early voice AI systems.

Modern streaming pipelines are very different.

Instead of waiting for each stage to finish completely, components can work simultaneously.

For example:

The system doesn't have to wait for 100% ASR → 100% LLM → 100% TTS before speaking.

A 2026 technical study on enterprise real-time voice agents found that the industry-standard approach remains a streaming STT → LLM → TTS pipeline and demonstrated a 947 ms median time-to-first-audio, with a best case of 729 ms, using cloud LLM APIs. (arXiv)

That is fast enough for many business conversations.

So the real question isn't: "How many components are in the architecture?"
It's: "How efficiently do those components work together?"

S2S Is Faster—But Faster Isn't Always Better

Native S2S has a genuine advantage in conversational responsiveness.

The architecture can reduce some intermediate processing and may provide better turn-taking and more natural speech interaction.

For example:

Customer: "I want to change my appointment..."

The AI can respond naturally without waiting for a traditional text pipeline to complete.

But call center performance isn't measured only by latency.

It is also measured by:

• Did the AI understand the customer?
• Did it follow the script?
• Did it call the correct API?
• Did it update the CRM?
• Did it follow compliance rules?
• Did it qualify the lead correctly?
• Did it transfer the customer at the right time?

And this is where cascaded systems remain very competitive.

Intelligence and Control Matter More Than "Human-Like"

A voice agent can sound incredibly human and still be a poor call center agent.

Imagine an AI voice agent that sounds perfect but:

• Gives the wrong price
• Books the wrong appointment
• Misunderstands a qualification rule
• Fails to update the CRM
• Doesn't follow the campaign script

That's a serious business problem.

Daily's 2026 benchmark is useful here. In its long, multi-turn enterprise-style evaluation, GPT-4.1 achieved 94.9% on the reported instruction-following metric, compared with 86.7% for GPT Realtime. Daily notes that the benchmark is only one data point and does not measure every aspect of voice quality, but it highlights the trade-off between speech-native interaction and mature text-model capabilities. (Daily)

The lesson is simple: A more natural voice does not automatically mean a more capable call center agent.

The Biggest Advantage: You Can Choose the Best Model for Each Job

This is where cascaded architecture becomes especially powerful.

Suppose you're building an AI outbound call center. You might choose:

You can change any one of these without rebuilding the entire system.

If a better ASR provider appears tomorrow: Replace ASR.
If a cheaper, smarter LLM becomes available: Replace LLM.
If a new TTS model produces a better local accent: Replace TTS.

This modularity is extremely valuable in an AI market where models change every few months.

Cascaded Architecture Is Also Easier to Debug

Imagine an AI agent tells a customer:

"Your appointment is confirmed for Tuesday."

But the appointment was actually booked for Wednesday. Where did the error occur?

With a cascaded architecture, you can inspect:

This level of observability is extremely important for enterprise applications. It helps engineering teams troubleshoot problems, improve prompts, monitor compliance, and continuously optimize the voice agent.

It Also Makes Compliance Easier

Contact centers operate in regulated environments. Depending on the industry and country, companies may need to manage:

• Call recording
• Consent
• Disclosure requirements
• Data privacy
• Customer verification
• Script compliance
• Sensitive information
• Human escalation

A cascaded system gives businesses clear intermediate representations. The conversation can be transcribed, classified, analyzed, and audited.

This doesn't automatically make a system compliant, but it provides a strong foundation for building compliance controls.

Cascaded Architecture Is Not Standing Still

Another reason not to underestimate cascaded voice architecture: It is evolving too.

Modern cascaded systems are becoming increasingly sophisticated.

Instead of:

we now see:

Everything can operate in parallel.

New research is also exploring full-duplex cascaded systems that allow the pipeline to preserve the modularity of ASR → LLM → TTS while improving interruption handling and conversational turn-taking. (arXiv)

This is important because it means the gap between the two architectures isn't necessarily: Old cascade vs. futuristic S2S.
It is increasingly: Optimized modular voice AI vs. end-to-end voice AI.

What About S2S?

This doesn't mean S2S isn't important. Quite the opposite. Speech-to-Speech is likely to become increasingly powerful as models improve.

It is particularly attractive for:

• Highly conversational customer service: Where natural interaction is critical.
• Virtual receptionists: Where the AI needs to feel like a human receptionist.
• Personal assistants: Where tone and conversational context matter.
• Voice-first applications: Where speech itself is the primary interface.
• Emotion-sensitive conversations: Where information beyond words may be important.

S2S will likely become a major part of the future voice AI landscape.

The question is when it becomes the better economic and operational choice for mainstream enterprise contact centers.

Why Cascaded Is Likely to Remain Mainstream in 2026

There are several reasons:

• 1. It already works: Companies don't need to wait for a new architecture.
• 2. It satisfies most call center requirements: Most contact center conversations are not philosophical conversations. They are things like:
  • "Where is my order?"
  • "I'd like to book an appointment."
  • "Can you tell me the price?"
  • "I'd like to cancel."
  • "Are you interested in this offer?"
  • "Can I reschedule?"
  • "What documents do I need?"
  These tasks can be handled extremely well by a properly designed cascaded system.
• 3. It costs less to optimize: Each component can be selected based on its cost/performance ratio.
• 4. It provides greater flexibility: Companies aren't locked into one speech model.
• 5. Text LLMs are extremely mature: The ecosystem around text-based LLMs—including prompting, RAG, tool calling, fine-tuning, evaluation, and observability—is extensive.
• 6. Enterprise integrations are easier to manage: CRM, ERP, APIs, knowledge bases, ticketing systems, and business workflows can be integrated around the LLM layer.

Daily's 2026 benchmark explicitly states that most production voice agents today use cascaded pipelines, citing their advantages in intelligence, observability, flexibility, and cost. (Daily)

A separate 2026 technical tutorial on building enterprise real-time voice agents also describes the cascaded streaming pipeline as the industry-standard approach in its study. (arXiv)

So while exact market-share numbers are difficult to establish, the available 2026 evidence supports a reasonable conclusion: Cascaded ASR → LLM → TTS remains a mainstream production architecture for enterprise voice agents in 2026.

The Future May Be Hybrid

The long-term answer probably isn't: Cascaded OR S2S.
It may be: Cascaded + S2S + specialized AI models.

This allows businesses to use the right technology for the right job.

And that is likely to be the more important trend in enterprise voice AI.

So, Which One Should Your Call Center Choose?

If you're building a production AI Voice Agent today, here's a practical rule of thumb.

The Real Question Isn't "Which Architecture Is Newer?"

The better question is: Which architecture delivers the best business outcome at production scale?

For a contact center, that usually means balancing:

• Voice Quality
• Latency
• AI Intelligence
• Integration
• Reliability
• Compliance
• Cost

And when all seven are considered together, cascaded architecture remains extremely competitive in 2026.

It may not be the most futuristic architecture. But that's exactly the point.

You don't need futuristic technology to solve today's call center problems. You need technology that works reliably, integrates with your business, scales to thousands or millions of calls, and keeps your cost under control.

Final Verdict

Native Speech-to-Speech is an important step forward for AI Voice Agent technology.

It can deliver impressive naturalness, faster interaction, and richer speech-level understanding.

But S2S has not made cascaded architecture obsolete.

For enterprise contact centers in 2026, ASR → LLM → TTS remains a highly practical—and likely mainstream—architecture because it combines mature technology, strong intelligence, flexibility, observability, and, importantly, better cost control. (Daily)

And for many call center applications, that's more than enough.

The future will undoubtedly bring better S2S models.

But until their economics, reliability, tool integration, and enterprise controls catch up across the board, the smartest approach may not be to chase the newest architecture. It may be to optimize the architecture that already works.