We are (RE)launching today!
🚀 Loqua is LIVE on Product Hunt today, and welcome to the family!
AI keeps getting smarter — yet using it feels more tiring: prompting, switching apps, screenshotting, explaining context again and again. Often the problem isn't the AI; it's that we still have to learn how to talk to it.
So we built more than better voice typing. Just say what you want, and Loqua understands your app, screen, language, and tone — dictate in Gmail, ask about a chart you're looking at, tell Codex or Claude Code what to change, set reminders, open apps, or have a page read aloud.
We believe voice agents are a necessary path to AGI — and a core part of physical AI. As a model-first team, Loqua is our first step: our own multimodal model for better transcription, translation, and multimodal Q&A, with deeper agentic, long-horizon abilities rolling out.
If this resonates, we'd love your support 👇
❤️ An upvote unlocks a promo code, and sharing earns extra days. Try it and tell us what you'd want it to do next.
👉 Upvote & try Loqua:
👉 [Website link]: https://theloqua.ai/
👉 [Product Hunt link]: https://www.producthunt.com/products/loqua?launch=loqua-2
Replies
Refocus
Congrats on the relaunch, Joshua. The detail that stands out is that you trained your own multimodal model instead of wrapping an off-the-shelf ASR, because that is exactly where the ceiling is.
I build voice AI for aging-in-place at Callie Care, where the calls go to people in their 70s and 80s, and the thing that broke every general-purpose ASR we tried was not accent but cadence: long mid-sentence pauses, hearing aid feedback, a TV on in the background. Endpointing fires early, the model talks over the person, and trust is gone.
Since you own the model: do you adapt endpointing per user over time, or is it a fixed threshold? And do you track word error rate on slow or atypical speech separately from your headline number? Curious how you think about that tail.
Loqua
@igorgurovich Very hands on question! Think of another way, traditional acoustic + LM feature won't fix this issue. The endpoint or thresholds tuning will only get you endless pain. We use semantic VAD and semantic token alongside with acoustic tokens in ASR to model cadence and user intention. To be more specific, the model knows if you finish the query or not, and acoustic events/ intentions/tool calls are all tokens on the run. This simplified the problem into a data driven problem. We track WER together with task finished rate to monitor E2E task completion status. However the challenge comes naturally with more labelled data (synthetic only for privacy).