For a child practicing English, an AI tutor needs to do more than mark an answer right or wrong. It needs to listen, understand the lesson, and respond quickly enough to keep the conversation going. EZtalking worked with SIRAYA to connect real-time voice, contextual AI, natural speech, and image creation within its existing learning platform.
Customer Overview
EZtalking is an AI education platform serving cram schools, teachers, and English learners in Taiwan. It brings speaking practice, AI-assisted pronunciation assessment, and digital materials into everyday teaching. Teachers can assign exercises and track progress, while children practice through games, visual prompts, and guided conversations in and beyond the classroom.
The goal is to make AI a learning companion that can follow a child’s question, respond in the context of the current lesson, and encourage another attempt—not simply judge an answer.
The Challenge: Make AI Responses Feel Natural
Children speak at different speeds, pronounce words differently, and often ask unexpected questions. To sustain a useful learning exchange, the platform must understand what a child says, connect it to the teaching material, and answer without a long pause. EZtalking’s earlier self-built language model could be slow to respond, and its synthesized voice could sound mechanical. The team also wanted voice experiences better suited to Taiwan, including Taiwanese Hokkien.
Visual content posed another constraint. Classroom exercises rely on pictures and scenarios, but limited local computing capacity meant some teaching images took days to prepare. That made it difficult to refresh materials as teachers’ needs changed.
The Solution: Connected Voice and Visual Learning
SIRAYA helped EZtalking connect the steps from a child’s spoken question to a relevant spoken answer. Agora RTC provides the low-latency audio foundation. EZtalking’s own APIs, language model, and internal retrieval-augmented knowledge base use the current teaching material to interpret the question and form a response. BytePlus TTS then turns that response into natural-sounding speech.
Together, these components help the AI stay within the lesson and keep the exchange moving. The voice workflow also supports speech experiences in Taiwanese Hokkien, making interaction more familiar for children in Taiwan. SIRAYA Model Router is being evaluated to give EZtalking more flexibility in choosing models for different teaching tasks; it is not included in the deployed outcomes described here.
For teaching materials, SIRAYA introduced Lumina cloud image generation to reduce dependence on limited local computing resources. Teachers can create images for vocabulary, lesson themes, and learning scenarios more quickly, giving children a visual context for the words they hear and speak.
The Outcomes: Faster Preparation, Better Interaction
Teaching images prepared in hours instead of days
Image creation that previously took days can now be completed in hours, helping teachers update vocabulary cards and scenario-based materials closer to the time they are needed.
More fluid AI speaking practice
Faster, more natural voice responses reduce the waiting and robotic feel of earlier interactions. Children can ask, listen, and try again in a more continuous practice loop, with Taiwanese Hokkien speech support adding local relevance.
A broader base for expansion
In parallel with the rollout of its AI teaching offering, EZtalking reports reaching 80% coverage of Taiwan’s cram school market. The connected voice and visual workflow gives the platform a foundation as it explores opportunities in other education markets, including Vietnam.