Voice agents should become more than transcription tools. Instead of simply turning speech into text, they should understand what users are trying to achieve. The same spoken idea may need to become a short Slack message, a detailed email, a technical issue, or a personal note. A useful voice agent should adapt automatically.
This is the direction Loqua is exploring: voice as an intelligent layer across the apps people already use. It can understand the current application, nearby content, personal vocabulary, and writing style, then shape spoken thoughts into something ready to use. Users should be able to speak naturally without organizing every sentence beforehand.
The next step is real-time duplex conversation. A good agent should know when someone has finished speaking, when they are only pausing, and when they want to interrupt. It should also recognize signals beyond words, such as emotion, hesitation, emphasis, and tone. Its own voice should feel equally natural, expressive, and appropriate to the situation.
Behind this conversation, the agent needs deeper capabilities. It should remember relevant context, think through difficult requests, use tools, and continue working after the conversation moves on. The voice experience remains fast and fluid, while more complex reasoning happens quietly in the background.
We all know that long-horizon coding tasks are difficult, but understanding long, multi-turn conversations is even harder. The agent must remain consistent, respond with the right emotion, use tools correctly, and remember what matters.
Dictation tools always mangle product names and internal jargon, which is half of what I'd be dictating. Is there a way to teach Loqua custom words, or does it pick them up from whatever's on screen? That part would decide it for me.
@protsenkoalexandra Totally get that, product names, acronyms, and internal jargon are exactly where generic dictation tends to get frustrating.
Loqua supports a personal dictionary, so you can add names and domain-specific terms you use a lot instead of hoping the model guesses them correctly every time. Screen context can also help when a term is visible, but we don’t rely on that alone.
If that’s the deciding factor for you, I’d definitely stress-test it with your real vocabulary, that’s exactly the kind of use case we want to get right.
@protsenkoalexandra We designed customized dictionary and a set of algorithms to auto pick your terms on screen for context. The more you use it, the better it works!
Great Product! But I have a question does Capture to Ask hold up on all screens like a spreadsheet or code editor, or is it mainly reliable on simpler layouts like the calendar and chat view shown here?
@sidraarifali It’s not limited to simple layouts, we’ve designed Capture to Ask for things like spreadsheets, dashboards, code editors, docs, and other dense screens too.
That said, the more complex the screen, the more important it is to capture the relevant area rather than everything at once. For a spreadsheet or code editor, that usually gives Loqua much cleaner context and a better answer.
So yes, those are definitely target use cases for us, not just calendar/chat-style screens.
@sidraarifali It works on all conditions, complex tasks such as spreadsheets, formulas, candlestick charts or picture caption, you name it!
The omni model choice is the part that stands out to me, since most stacks still drop the screen context somewhere between ASR and the LLM. I work on voice AI for daily check-in calls with older adults, and the piece that never gets easier is endpointing: telling a real end of thought from a mid-sentence pause, which gets harder the more rambling the speech is. How does Loqua decide you are finished talking during a long dictation? And does the screen context feed into the rewrite step, or only into command routing?
@igorgurovich Great question! Endpointing is one of the areas where we really didn’t want to rely on a fixed silence threshold.
Loqua uses semantic VAD and semantic tokens alongside acoustic tokens, so the model is looking at cadence and intent, not just “how long has the user been quiet?” The goal is to distinguish a genuine end of thought from a mid-sentence pause, especially in longer or more meandering speech.
And yes, screen context can feed into the rewrite/understanding step too, not just command routing. We want the model to use what you’re looking at as part of understanding what you mean, rather than treating screen context as a separate routing signal.
Your use case with older adults sounds especially challenging btw: slow cadence, long pauses, background audio, and barge-in are exactly the kinds of cases where naive endpointing falls apart.
Screenshot to ask is such a simple idea but removes so much friction. No more explaining screenshots manually.
@zaniaz Yep. Just capture what you're looking at and ask about it directly.
How long does setup actually take? I don't want to spend 20 minutes configuring another AI tool.
@renchu_song It's designed to be pretty lightweight: install, sign in, grant the required permissions, and you're ready to go.
The screen context part is what caught my attention. Being able to just point at something and ask instead of explaining what I’m looking at every time sounds genuinely useful. Curious to try this in my daily workflow!
@mi_leeli Exactly, that’s one of the interactions we really wanted to make feel effortless. Sometimes explaining what you’re looking at takes longer than the actual question 😅
Really curious to hear how it fits into your day-to-day once you’ve tried it!