Playing Around with Amazon Nova 2.5 Sonic

Playing Around with Amazon Nova 2.5 Sonic

A new version of Nova Sonic came out without my noticing. How good is it, really..?
2026.11.22

This page has been translated by machine translation. View original

Introduction

Do you talk with AI every day?
What do you usually use when you talk?
Among the people around me, I feel that text (human)↔︎text (AI) or voice (human)↔︎text (AI) are the most common.
That's true for me too.

However, depending on the situation, there are times when I want to interact with voice (human)↔︎voice (AI).
Recently, the ChatGPT and Gemini smartphone apps have been adding more of these voice-to-voice features for talking with AI, haven't they?

I'm talking about this type of AI feature that our company's Aramaki wrote about in a blog a little while ago.
https://dev.classmethod.jp/articles/gpt-live-vibe-cooking/
This blog is interesting, so I recommend it.

I want to put this kind of feature into my own product too!
Actually, AWS also has this kind of voice model.
Amazon Nova Sonic. You've all heard of it, right????

Amazon Nova Sonic?

Amazon Nova Sonic reached GA in 2025 as a Speech-to-Speech model.
https://aws.amazon.com/jp/about-aws/whats-new/2025/04/amazon-nova-sonic-speech-to-speech-conversations-bedrock/

I also touched it a little when it first went GA, but it felt, how should I put it... like talking with someone in a different time zone with a delay of a few seconds.
I kept my distance from it for a while after that, but I only now learned that last month a new version called Nova 2.5 Sonic (hereafter written as Nova Sonic) was made generally available.
https://aws.amazon.com/jp/about-aws/whats-new/2026/10/amazon-nova-2.5-Sonic/

Trying it out

Well, I figured I'd try it out for now,
so I tried Nova 2.5 Sonic in the Bedrock playground.
CleanShot 2026-11-20 at 23.32.36@2x
I was amazed.
It's as smooth as talking with a human!!!!!!

I want you, who are reading this, to open the AWS console right now and try it too.
It's easy if you just open the playground and try it.
If you have tried the previous Nova Sonic, I think you'll be surprised.

There are occasionally slightly odd bits of Japanese, but it's at a level that barely bothers me.
To begin with, Japanese isn't listed as a supported language for Nova Sonic in the official documentation, so I'm grateful that it even responds in Japanese.

Also, even when I interrupted in the middle of the conversation, it responded nicely, so I feel the exchanges have become much more natural.

On the other hand, for somewhat difficult questions it sometimes replied "I can't answer."
It can't answer everything, but the tempo of the conversation was as natural as talking with a human.

Thinking about how to use it

While models like this are fun, there are times when it ends with just "Wow, amazing. clap clap"
So I thought about how to put it to use in my own way.

We engineers probably communicate with AI Agents via text as naturally as breathing in our daily lives.
Honestly, I don't think it would fit that group very well.

Rather, I think it would be better to use it by embedding it in applications for people who don't usually spend much time facing a computer.

In my previous job, I was a mechanical engineer, so I only used a computer for clerical work, and I spent more time dealing with facility equipment otherwise.
I wasn't very good with computers, so now I think, "I'd have been happy if there had been an AI back then that did clerical work while interacting by voice," or "I'd have been happy if I could ask an AI various things conversationally and have it teach me."
Pasted image 20261121224133
Beyond that, I imagine it could also be used to have an AI interview people who rarely face a computer but have specialized knowledge, getting them to put into words the tips for judgment and rules of thumb that are not normally written down in documents.

Nowadays there is a desire to turn tacit knowledge into data and make use of it,
but for my past self, who rarely touched a computer, even if someone had said "We want to turn what you know into data, so please write it out in text," I think it would have been quite difficult to put it into writing.
Pasted image 20261121223854

I'm hoping voice models can also be used for drawing out that kind of specialized knowledge, which doesn't get turned into data, from people.
So, I played around with Nova Sonic to find out what it can do.

Talking with a voice conversation app

The app itself was built by asking Kiro to "build an app with Nova 2.5 Sonic."
It builds things quickly, so if you want to try it, ask the AI Agent at hand.

There were a few things I wanted to verify, so the configuration is like this.
This time I built an Agent that helps with internal applications by voice.
architecture-v2 (3)
Setting the details aside, let's first try how it behaves.

Demo

The main things I'm testing in the video are these two:
• Whether it can create an application form while answering questions
• Whether it answers by referring to internal data

https://www.youtube.com/watch?v=-Ao3hjqB-Cs

Nice, isn't it?
There are some parts where the Japanese is slightly off, but my answers are immediately reflected in the sheet.

It can also answer things that are written only in internal documents or internal databases as needed. (Of course, it's dummy data.)

For the question I asked midway, "How many days of paid leave do I have left?", a Lambda called via MCP retrieves the data from DynamoDB and calculates it as "paid leave days at the start of the period − days used."

Also, for the question "Can I take paid leave in half-day units?", it retrieves the internal regulations document via MCP.
I had prepared many sample internal documents, so there seems to have been a pause for the search, but it's not bothersome.
I'm curious how much lag would occur if the number of documents grew much larger.
Pasted image 20261122130133
This time I built an internal application Agent to keep things easy to understand, but if possible, I'd like the AI to ask questions from various angles and draw out experiences and judgment tips that are hard to summarize in text.
Then transcribe that conversation, organize and structure it with AI, and finally have the person confirm it.
I thought that kind of use would be interesting.

This is my impression after trying a voice model this time:
(Human→AI) It's fine for a human to ask the AI questions, but
(AI→Human) having the AI ask people all sorts of questions and get them to let out things that are tedious or difficult to write as text,
and then using another AI to shape that into data. I think this kind of use could differentiate it from text-based AI Agents.

Key points of the configuration

Have it ask questions following the items of the application form

If you vaguely just ask "I want to apply for paid leave," it may ask unexpected questions, so this time I created the prompt so that it loads the appropriate internal application form (dummy) based on the user's request and asks only questions that follow the items in the document.
leave-form-flow

I feel that, unlike text, spoken conversation isn't well suited to deep deliberation.
So this time I took the approach of having the AI come up with set questions based on existing documents.

Calling Nova Sonic from Bidi Agents

For the Agent that calls Nova Sonic, I used an Agent called Bidi Agents.
To talk with AI in a human-like conversation, you need bidirectional real-time streaming, right?
Bidi Agents is one of the libraries included in the Strands Agents SDK, a bidirectional agent for real-time streaming conversations.
It also supports Nova Sonic, which we use this time.

The configuration this time runs Bidi Agents on AgentCore Runtime and calls Nova Sonic.
Here's the picture.
bidi-agentcore-nova-sonic
AgentCore Runtime can stream audio bidirectionally over WebSocket, so it's also a good match as a platform for running Bidi Agents.
Also, if needed, it's easy to integrate with AgentCore Gateway, which manages the MCP tools described later.

Connect MCP tools to include company data in answers

Nova Sonic alone is fine, but you'd really want it to hold conversations that take into account information that only your company or you have, in addition to information available in the world.
So in this configuration, I placed dummy company data in S3 and DynamoDB.
The Lambda that calls the company data is registered in AgentCore Gateway as an MCP tool.
bidi-agentcore-gateway-mcp
With this, when Nova Sonic requests a call to a necessary tool, BidiAgent executes the Lambda via AgentCore Gateway and can answer using the retrieved company data.
In the demo too, for the question "Can I take paid leave even for half a day?", it answered based on the internal regulations.

Conclusion

I wrote that this is for non-engineers, but I think many people find it difficult in the first place to sit alone at a PC in the office and talk.
Also, quietly talking alone with a voice AI feels a bit odd.

I get the impression that voice models are hard to find uses for, but I think conversation is a treasure trove of valuable information that gets left out of text communication, so I'd like to explore good ways of using it that fit into our work.

If you have any interesting ways of using it, please let me know.

Share this article

AWSのお困り事はクラスメソッドへ