Loka builds a sub-second-latency voice agent on Amazon Nova 2 Sonic

Loka has published a detailed breakdown of the architecture behind its voice agent built on Amazon Nova 2 Sonic, AWS's next-generation speech model. The goal was specific: build a bot that customers won't hang up on after a few seconds of waiting.

The problem they set out to solve

The robotic voice of phone bots is not just an aesthetic irritation. For businesses, it means direct losses: the customer hangs up, calls back to reach a live operator, or switches to a competitor. Brand reputation suffers and support costs climb.

Classic voice systems run through a long pipeline: speech recognition (ASR) → conversion to text → language model → response generation → voice synthesis (TTS). Latency piles up at every stage. The result is a gap of 2 to 5 seconds between the customer's question and the bot's answer. In that time, the caller decides the system isn't working and either hangs up or demands a live operator.

Loka set out to break this chain and build an agent that responds within a natural conversational pause, the way a human would. Its answer was Amazon Nova 2 Sonic.

What Nova 2 Sonic does differently

Nova 2 Sonic is a multimodal speech-to-speech model from AWS that works directly with audio and skips the separate ASR transcription and TTS synthesis steps. It takes an audio stream as input and produces an audio stream as output, with no intermediate conversion to text.

That fundamentally changes the latency profile:

• Responses begin within 300–500 ms after the user pauses
• The model understands natural interruptions in speech and handles them correctly
• The system hears intonation and emotional context, and adapts the tone of its reply
• The "system is processing" feeling disappears from the conversation entirely
• Integration with business logic via function calling doesn't break the flow of the conversation

Nova 2 Sonic is available through Amazon Bedrock, so companies already on AWS can plug it in without switching providers or rebuilding their infrastructure from scratch.

The production architecture

Loka streams audio in real time with minimal buffering. The system doesn't wait for the user to finish speaking. It starts processing immediately, which lets Nova 2 Sonic respond right at the natural pause rather than after a drawn-out silence.

"A robotic voice is the main reason customers hang up.

This isn't a technology problem — it's a trust problem," the Loka team notes.

To pull business data in real time, such as order status, customer history, and inventory levels, the agent uses real-time function calling. To the customer, this feels like an instant answer rather than a noticeable "waiting for results" pause.

In production, the system holds up against interruptions, topic changes, and unusual pauses, the scenarios where classic ASR systems most often fail.

What it means

Speech-to-speech models remove the main barrier to mass adoption of voice bots: the noticeable latency that shatters the illusion of a live conversation. If the latency goes unnoticed and the voice sounds natural, the line between agent and operator blurs.

For businesses, this is a direct path to automating the call center without taking a hit on NPS. Other providers will follow Nova 2 Sonic with similar models, and competition in the voice AI segment is only getting started.