The client is a leading consumer-electronics brand that builds and supplies low-cost PCs, smartphones, tablets, and wearables, present in more than 70 countries and well regarded in emerging markets across Europe, the Middle East, Africa, and the CIS region. They had developed a version of their Android tablet designed specifically for children, and wanted to add a voice assistant - a kid-friendly way for children to interact with the tablet by voice, triggered by a wake word and working both online and offline.
Building a voice assistant for children, on low-cost hardware, that works offline and understands many English accents, is a tough combination — every requirement pushes against the constraints of the device.
The tablets were low-cost, with modest hardware that lacked high-end performance - a real constraint for on-device speech processing.
The client had several tablet versions with different hardware configurations and custom Android builds, complicating a single solution.
The app had to be integrated with and locked to the target hardware so it would not run on any other device.
It had to work offline against a question library and fall back to online search, while recognising a range of global English accents and stay engaging and intelligible for children.
Focaloid set up an agile product team and built a kid-friendly voice assistant on NLP, ASR, and TTS - answering from an offline library, falling back online, and tuned to children and to constrained hardware.
Assembled a team of architect, tech lead, backend developers, Android developers, and QA analysts, working in bi-weekly sprints to iterate quickly on feedback.
Built on Natural Language Processing (NLP), Automatic Speech Recognition (ASR), and Text-to-Speech (TTS) converting a child’s speech to text, processing it, and playing back the answer as voice.
Defined wake words that initiate the assistant, chosen after comparing recognition across different accents.
Answered from a predefined offline question library, using the Google Assistant API as a fallback for questions not in the library when online.
Used a kid’s-voice TTS and face emoticons for visual feedback, with intelligent follow-up probing when a question isn’t recognised, and a question log to capture unanswered questions for future library updates.
Focaloid built iteratively in bi-weekly sprints, tuning for children, accents, and constrained hardware.
Demonstrated continuous progress and iterated quickly based on client feedback.
Implemented the ASR → NLP → TTS pipeline and selected wake words validated across accents.
Chose a kid’s-voice TTS from sampled voices, added emoji feedback, and fine-tuned performance for low-cost hardware.
Built the offline question library and the Google Assistant API fallback, plus a question log to drive future library updates.
A voice assistant for children has a higher bar than most: it has to understand small voices and many accents, answer reliably even with no internet, and feel friendly and engaging all on inexpensive hardware that wasn’t built for heavy speech processing. Getting all of that right is what makes the difference between a gimmick and a feature kids actually use. By building an offline-first, kid-friendly assistant tuned to constrained devices with a child’s voice, emoji feedback, accent-aware wake words, and sub-2-second offline answers - Focaloid gave the client a genuinely usable voice experience for its children’s tablets, ready to roll out across more devices.
We build NLP, ASR, and TTS-powered voice assistants tuned for your hardware and your users - offline-capable, accent-aware, and engaging.