In my previous posts, Self-Healing Tests with AI and AI Coding Agents in 2026, I tried to strongly emphasise an important new capability offered by LLMs such as Fable and the latest Opus. These models can support discovery and learning. They allow us to have relatively abstract conversations, perform research, explore possible solutions, look for different proposals, and ultimately determine what should actually be done.
This kind of work is particularly well suited to voice. Discovery rarely begins with a perfectly structured prompt. It is usually much more exploratory. We need to talk through the problem, consider different perspectives, explain what we already know, and gradually discover what information is still missing.
It is also quite important to spend some time talking to the agent and giving it context, because this usually leads to better results. Voice makes that easier. Speaking allows us to provide more background, describe the problem more naturally, and include details that we might omit if we had to type everything.
What I would like to recommend and describe today is therefore the voice interface. In my view, this interface offers several important benefits, which I will try to explain in this post.
One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.
Transfer More Context, Faster
One perhaps less obvious benefit of the latest LLMs is their ability to use accumulated knowledge selectively and sensibly. We could say that these models need fewer and fewer low-level instructions such as: do this, then do that, and finally do something else. They still, of course, need context. They need information, knowledge, and pointers towards anything that might help them solve a particular challenge.
This is precisely where the voice interface offers an advantage. We can usually talk for quite a long time. This is particularly true for people working in IT companies and for those who have experience with training or public speaking. Speaking is relatively fast, which is also worth remembering. It is less tiring than typing and, in most cases, allows us to communicate more information and provide more context.
By speaking, we can transfer our way of thinking. We can describe where we started, which problem we are trying to solve, what conclusions we have reached, what we have already checked, and what we have not checked yet. We can explain where we suspect the problem might be, where we see an opportunity for improvement, what we would like to achieve, and what we would prefer to avoid. We can also describe what would make us conclude that the result is good and what would tell us that it is wrong.
This gives us a powerful way to transfer a large amount of information very quickly and, as a result, practise highly effective context engineering. Voice allows us to move raw information into the model's context faster. This matters because the latest models are very good at working with unstructured knowledge. They can take these raw pieces of information, put them together, and organise them into something coherent. The information does not need to arrive as an elegant, perfectly structured prompt. We can provide the material, and the model can help us make sense of it.
Think in Your Native Language
Another, perhaps less obvious, advantage is that non-native English speakers can express their thoughts in their native language. English is, of course, the language of IT and the international language we normally use to communicate. I suspect, however, that many of my readers are not native English speakers. I know that I have many readers from Poland, because I am known primarily within the Polish software-testing community.
When I use a voice interface, I usually speak Polish. In fact, I am dictating this post in Polish right now. At the beginning of this session, I asked Codex to translate my thoughts into English as literally as possible while preserving my style. This is a significant benefit because it allows me to express myself in Polish while still producing an English article.
As I write this, I also have two other agents working. I asked one of them to add monitoring to my production LocalStack environment. The other is helping me create a neural-networks course. I communicate with both of those agents in Polish as well. At the beginning of each session, I ask the agent to respond in English, but explain that I will communicate in Polish.
I do this because Polish is the language in which I think. It is the language I grew up with and the language that allows me to communicate more nuanced knowledge. Even if my English were at a somewhat higher level, I believe I would still be able to communicate more information, faster and more accurately, in my native language than in a language I learned later. Polish is a language I know from the inside. I have used it since childhood, I think in it, and it is simply my native language.
I am aware that writing a prompt in Polish requires slightly more input tokens. In the pricing models I currently use, however, output tokens are more expensive than input tokens. I therefore do not consider this a significant waste of money. It is a cost worth accepting because the benefit of providing better context is greater than the additional cost of prompting in Polish.
Voice Is Becoming a First-Class Interface
The voice interface is already highly developed in leading AI products. As I write this, I am using Codex. I subscribe to Codex and ChatGPT, and both the conversational ChatGPT Voice experience and voice dictation available when working with Codex are excellent.
I speak Polish, and the transcription is remarkably accurate. Even if I mumble slightly, what I say is transcribed correctly and translated into English without any problems. This makes voice feel like a normal and dependable input method rather than an experimental addition.
Anthropic is investing in the same direction. Claude has a dedicated Voice Mode, which supports complete spoken conversations in its mobile applications. Cursor also supports voice input for working with agents and has continued to improve the quality of its built-in speech-to-text transcription.
We can therefore say that voice is becoming a first-class interface across the leading AI tools. In my view, its importance will continue to grow.
Local Voice Is Already Practical
The voice interface is also becoming increasingly capable on local computers. I use Freestyle, which is a free and open-source dictation application, together with its Whisper Pro option. Underneath, it uses OpenAI's multilingual Whisper large-v3-turbo model and performs the transcription locally.
The model occupies approximately 1.6 GB on disk and Freestyle recommends around 6 GB of RAM. These requirements are not particularly high for a modern development machine, especially when we compare them with the resources required by local language or image-generation models.
This setup works very well for me. I can speak Polish, dictate long prompts, and receive accurate text without sending the recording to an external service or paying for API usage. I am very satisfied with the results and can recommend Freestyle with Whisper Pro to anyone who would like to experiment with a local voice interface.
Conclusion
Using voice is challenging because it requires us to change some of our habits and established ways of working. It is not something that comes easily, at least in my experience. It requires us to develop new habits, and, as we all know, habits are very difficult to change.
From my perspective, however, I can strongly recommend making the effort. Since I started talking to my computer, I have noticed that I simply work better with agents. I provide more context, express my thoughts more naturally, and get better results.
And, at least so far, I have not gone mad. 🙂
Comments
Loading comments...