EchoRead™ SDK · K–3 reading
An on-device speech intelligence platform for early education.
It listens on the device to how a child reads, not just what they say, then scores fluency, coaches out loud and adapts the level.
Simplified web demo · Reading level 3.1
The small fox ran to the red barn.
Try skipping a word.
Your voice stays in this tab.
This web demo runs an open-source speech model, not Wundr’s models for children’s voices. EchoRead runs inside your app, on the device. Wundr reading levels, 1.1 to 4.4. The demo downloads 40 MB once.
01 / Product
The problem
Parents and teachers:
stretched thin
Parents are short on time. Classrooms are crowded. 1-1 reading practice is rare.
Tutors: hard to afford,
hard to find
Many families can't afford a tutor. Good tutors rarely have hours that fit a family's schedule.
Every child reads at a different level, and at a different pace.
Fluency grows by reading aloud to someone who listens.
Why now. Most speech recognition was built for adult voices and struggles with children's speech. We build streaming speech models for children's voices that run on everyday devices.
Rooted in the Science of Reading.
Skilled reading occurs when word recognition and language comprehension become fluid and automatic. EchoRead™ bridges both strands, coaching decoding and vocabulary during oral reading practice while continuously evaluating oral reading fluency.
Adapted from Scarborough’s Reading Rope (2001).
Listens as the child reads. Hints when they’re stuck.
Listening
Matches what a child says to the words on the screen, word by word, in real time, and detects miscues.
Coaching
Gives a gentle hint the moment a child gets stuck, then moves them to harder text as they progress.
Pronunciation & Vocabulary
When a child stumbles, a word card breaks the word into parts and sounds, says it, and asks the child to try again. Word sets adapt to each child’s reading level.2
Built-in assessment and progress
Progress captured automatically. Real-time insights built into every reading session.
Turn everyday oral reading into formative data: accuracy, words correct per minute, and every miscue by type, with the word the child actually said. Text adapts to each child’s level as they read. Results stay on the device, and your app decides what parents and teachers see.
Built for how children actually speak.
1. More than two in three U.S. fourth graders (69%) did not reach NAEP Proficient in reading in 2024. Source: The Nation's Report Card, 2024 NAEP Reading Assessment.
2. Word sets are leveled using Dale-Chall readability and Age of Acquisition data.
02 / Privacy
Private by design
A child's voice never leaves the device.
Children's products face strict privacy laws like COPPA and FERPA. EchoRead™ keeps a child's voice and reading data on the device, so there is no voice data on a server to secure or audit.
Read the privacy noticeFor a six-year-old, voice isn't the primary interface. It's the only interface.
03 / Technology
One SDK, inside your app.
For early education, starting with reading. EchoRead™ is a platform (SDK), not an app: bring your own reading content or leveled curriculum, and EchoRead adds the speech and oral reading fluency, comprehension and assessment layer to your education product. It runs on the device, at the edge.
* Only the coach’s sanitized reply text goes to the cloud Text-to-Speech service. A child’s voice and reading data never leave the device.
† Mainly Wundr’s own recordings, collected with verifiable parental consent.
Follows the reader live
The highlight tracks each word as the child reads, typically under 200 milliseconds.
Two models must agree
Two independent on-device low-latency speech models check every read. Echo only corrects a child when both agree.
Speech-to-text turns speech into text. The Streaming Reading Engine starts with the text, and listens for the speech.
It transcribes too, so you still get the words, and it tells you where a child’s reading departs from the page.
| Feature / metric | Wundr on-device SDK | Traditional cloud voice APIs |
|---|---|---|
| Latency | Real-time word tracking (typically under 200 ms), on the device | Network-dependent (adds a round-trip delay) |
| Data privacy & compliance | Child audio stays on the device. Designed to support COPPA & FERPA | Complex (audio sent to the cloud) |
| Per-student compute cost | No per-minute cloud or token fees for listening and scoring | Metered per minute or per token |
| Classroom Wi-Fi reliability | Listening and scoring work offline | Slows or stalls on weak school Wi-Fi |
| Literacy focus | Streaming reading engine, plus patented oral reading fluency evaluation and adaptive feedback | Generic speech-to-text |
Listening, scoring and coaching run on the device, on iOS and Android.
04 / About
Why we built EchoRead™.
Learning to read aloud depends on someone listening.
Wundr AI builds on-device speech intelligence that runs directly on the hardware children already hold. We started with early literacy because a child can only practice reading aloud when someone is there to hear them. EchoRead™ adds an intelligent, private listening layer into the learning apps kids use every day.
The founders
Reno Marioni
Co-founder and CEO
Reno is an AI builder, product innovator and two-time founder, with executive product roles at Nokia, Uber/Careem, BMW and HERE. Originally from San Francisco and now based in Miami, he leads Wundr's product, patents and business, and stays hands-on with the systems and the data behind our models.
As a parent watching his daughter navigate early literacy, he built EchoRead™ to give every child a patient, zero-judgment listener right on their device.
Ruben Gonçalves
Co-founder and CTO
Ruben is a full-stack AI engineer and systems architect with a Master's in Mobile Computing and a decade of experience across mobile infrastructure, developer tooling, and platform engineering.
Based in Portugal, he specializes in speech and audio AI and leads the technical architecture of Wundr’s on-device voice AI engine and platform SDK. Outside of software, he is an avid builder who designs and crafts custom motorcycles.
Add speech intelligence to your product.
Tell us what you are building. Early SDK access is by request, from a work email.