Speech Screening Prototype
Need to talk?

KERICE 2026International Islamic University Malaysia

Gold Award, KERICE 2026

Listen Early. Refer Sooner.

A Low-Cost Edge Device for Speech-Based Depression Screening

A small device that asks a few simple questions, listens to how a person speaks, and gives health workers an early screening indication. It runs offline, and every recording stays on the device.

  • PRGS25-029-0073

We also released pyvoicebox, an open-source Python port of the VOICEBOX speech toolbox

A screening aid, not a diagnosis. The decision always stays with the clinician.

Videos

Watch the Film, Then See It Running

Film · 3:09

The Project in Three Minutes

Why we built it, how it listens, and where it goes next. Subtitles are built in.

Live demo

See the Prototype Running

The device running a voice check-in from start to finish.

The demo plays from YouTube in privacy-enhanced mode; nothing loads from YouTube until you press play. Open it on YouTube

Read the film’s transcript

How can you tell when someone is quietly struggling? Around one in twenty adults lives with depression. Even in high-income countries, only about one in three people with depression receive treatment.

In Malaysia's rural and B40 communities, stigma keeps many people from asking for help. Community clinics rarely have the time or the staff to screen everyone. Most tools that could help also run in the cloud, far from where coverage is poor.

So we asked a different question. What if there were a way to spot the signs of depression from the voice alone, on a portable, low-cost device?

This is our prototype. A Raspberry Pi 5 with a small touchscreen and an external microphone, built for under three hundred dollars.

Here's how it works. The person answers a few everyday questions out loud. As they speak, it measures two things. The timing of speech: how long each answer takes to begin, and how often they pause. And the voice itself: how far the pitch moves, and how steadily it is sounded. Depression often shows as slower replies, longer pauses and a flatter voice.

Everything runs on the device itself. It needs no internet, and no recording ever leaves the unit.

Then it shows a plain-language result, and why. Its strongest cues are the slowest replies, how steadily the voice is sounded, and how far the pitch moves. Because it listens to the sound of speech, never the words, it is designed to work across languages.

Our model file is just 3.6 kilobytes, about a thousand times smaller than one phone photo. Because it is so small, it runs on a low-cost edge device, with no graphics card and no cloud.

Tested on recorded clinical interviews, it is competitive with published speech systems, some of them thousands of times larger.

It gives health workers an early indication of who may benefit from being seen. The decision always stays with the clinician.

We plan to make it free for rural and community clinics, NGOs and researchers, with an enterprise version for private healthcare. Its screening engine can also be licensed to digital health and workplace wellness platforms. And at under three hundred dollars, a single sponsor can put one in a clinic that has none.

Our next phase is to test and validate it with Malay speakers, in a clinical pilot at IIUM Kuantan. Listen early. Refer sooner.

Award

Gold Award at KERICE 2026

On 7 October 2026 the prototype received a Gold Award at KERICE 2026, the Kulliyyah of Engineering Research, Innovation and Commercialisation Exhibition at IIUM.

Thank you

To the Kulliyyah of Engineering and the KERICE 2026 organising committee for running the exhibition, and to the Dean, Assoc. Prof. Dr. Khairul Azami Sidek. To the judges, for their time and their questions. To the sponsors, including ChemBio, for supporting the event. To the Ministry of Higher Education Malaysia, whose Prototype Development Research Grant Scheme funds this work. And to everyone who visited our booth.

Gold Award certificate presented to Muhammad Fahreza Alghifari, Teddy Surya Gunawan, Mira Kartiwi, Ali Sophian and Jamalludin Ab Rahman for the exhibition of the Low-Cost Edge-AI Prototype for Speech-Based Depression Screening in Community Healthcare at KERICE 2026, 7 October 2026, signed by the Dean of the Kulliyyah of Engineering, IIUM.

The problem

Depression Often Goes Unnoticed

Hard to reach

In Malaysia's B40 households (the lower-income 40 per cent) and in rural communities, depression often goes unnoticed and people are not referred for care.

Stigma

Stigma keeps many people from asking for help. A quiet, private first step can make it easier to start the conversation.

Recordings in the cloud

Most tools that listen for signs of depression run in the cloud. That puts sensitive recordings on someone else's server, and rules them out where the internet connection is poor.

What is missing is a screening aid that is small, offline and private, so people can trust it where stigma is high.

How it works

It Listens to How People Speak, Not What They Say

Depression often shows in speech as slower replies, longer pauses and a flatter voice. The device measures those patterns from the sound alone.

One question and answer, drawn as soundThe device asks a question. About one second later the person starts to answer. Their answer contains a longer pause before they continue speaking.Device asksPerson answers1.0 sbefore the answer startspause0 s5 s10 s15 s
Illustration, not a real recording.
  1. 1

    The device asks a question

    Each question appears on the screen and is read aloud. No interviewer is needed to run a session, and because the device asks each question itself, it knows exactly when the question ends.

  2. 2

    It notices when the answer starts

    A small voice detector (Silero) hears the moment the person starts to reply. The gap in between is their time before answering.

  3. 3

    It measures how the person speaks

    Pauses, how far the pitch rises and falls, and how steadily the voice is sounded. It listens to the sound of speech and never to the words, so it is designed to work across languages.

From speech measures to a screening indicationFour speech measures, slowest replies, usual time before answering, pitch range and voice steadiness, feed a lightweight classifier with 33 parameters. It gives a screening indication, here follow-up suggested, for referral to a clinician. Everything stays on the device.Everything stays on the deviceSlowest replies2.8 sUsual time before answering1.03 sPitch range173 HzVoice steadiness0.65Lightweight classifier33 parametersScreening indicationFollow-up suggestedfor referral to a clinicianFrom speech measures to a screening indicationFour speech measures, slowest replies, usual time before answering, pitch range and voice steadiness, feed a lightweight classifier with 33 parameters. It gives a screening indication, here follow-up suggested, for referral to a clinician. Everything stays on the device.Everything stays on the deviceSlowest replies2.8 sUsual reply time1.03 sPitch range173 HzVoice steadiness0.65Lightweight classifier · 33 parametersScreening indicationFollow-up suggestedfor referral to a clinician
4

A lightweight classifier combines them

A 33-parameter classifier turns the measurements into a screening indication, to help decide who may benefit from being seen. The decision always stays with the clinician.

5

Nothing leaves the device

No internet connection is needed. Recordings are processed on the device itself and are never sent to a server.

The prototype

A Screening Aid That Fits on a Clinic Desk

A Raspberry Pi 5 (8 GB) with a 3.5-inch touchscreen and a USB table microphone. Built from off-the-shelf parts for under $300. The device asks each question on its own, so no interviewer is needed to run a session.

Inside

Everything Happens on the Board

The voice detector, the measurements and the classifier all run on the Raspberry Pi 5. No cloud server, no internet connection. Recordings never leave the unit.

  • Case
  • 3.5-inch touchscreen
  • Raspberry Pi 5 (8 GB)
  • USB table microphone

Raspberry Pi OS (Debian 13), Python 3.13.

On the screen

What the Device Shows

The device runs a short voice check-in on its own touchscreen. It asks each question aloud, shows what it hears as the person answers, then explains the result with the timing and voice measures behind it, each compared with a typical range.

  1. Kiosk screen titled Voice check-in: answer a few everyday questions out loud and watch how the device listens to the way you speak. About 5 minutes. Your voice is not kept. A Start button.

    1A short voice check-in that anyone can start from the screen.

  2. Kiosk screen showing question 1 of 8, How are you doing today?, with a live trace of the answer: when the person is speaking, their pitch, and a 1.0 second time before answering. Stop and Next buttons.

    2Each question is read aloud. As the person answers, the screen shows when they speak, their pitch and their time before answering.

  3. Kiosk screen titled Checking your voice, with the measurements so far: typical response time 0.9 seconds, time speaking 1 minute 15 seconds, answers heard 8 of 8. Your recording is deleted as soon as this finishes.

    3The voice is checked on the device in a few seconds, and the recording is deleted straight after.

  4. Kiosk result screen: No strong signs showed up. That is good to hear. If things feel hard, talking to someone can still help. Beside it, why this result: slowest replies 1.1 seconds, usual time before answering 0.89 seconds, pitch range 115 hertz and voice steadiness 0.70, each against a typical range. This is not a diagnosis, with support line numbers.

    4The result in plain words, with the four measures it is based on. It is not a diagnosis.

  5. Kiosk screen titled What it measured: slowest replies 1.1 seconds, shorter than typical; usual time before answering 0.89 seconds, within the typical range; pitch range 115 hertz, narrower than typical; voice steadiness 0.70, higher than typical. Each is drawn against the typical range.

    5Each measure compared with the typical range.

  6. Kiosk screen titled Your answers, one by one: a bar chart of the time before answering each of the 8 questions against a typical line, the pitch movement over the answers against a typical range, and the whole session with the device asking and the person answering.

    6The answers one by one: time before answering, pitch movement and the whole session.

Screens from the current kiosk app at the panel's real size (480 × 320), from a real check-in run on the device with its real model. The answers were spoken by a synthetic test voice, not a real person.

Read the result screen as text

No strong signs showed up. That is good to hear. If things feel hard, talking to someone can still help.

Slowest replies
1.1 s (shorter than typical)
Usual time before answering
0.89 s (within the typical range)
Pitch range
115 Hz (narrower than typical)
Voice steadiness
0.70 (higher than typical)

This is not a diagnosis. The screen also lists people to talk to.

When the patterns differ from the typical range, the result reads "Some signs worth talking about" and suggests talking to someone the person trusts, or a counsellor. If it heard too little speech, it says so and offers to try again.

Key numbers

Small, Fast and Private

11.7 s

to score an 11-minute interview

on the Raspberry Pi 5, about 55 times faster than real time, once the interview is recorded.

under $300

of hardware

Off-the-shelf parts: a Raspberry Pi 5, a small touchscreen and a table microphone.

33

-parameter classifier

Its whole model file is 3.6 KB. No graphics card and no cloud needed.

0.74

AUROC, cross-validated

How well the scores separate people with and without depression symptoms: 0.5 is a coin toss, 1.0 is perfect. Measured on 189 participants.

  • 9.0 s of the 11.7 s is voice-activity detection, a separate small neural network (Silero) that finds when someone is speaking.
  • 602 MB peak memory on the device.
  • Tested on DAIC-WOZ clinical interviews: 189 participants, counted as depressed at a PHQ-8 questionnaire score of 10 or more.

Results

In the Range of Published Speech-Only Systems

On the same clinical interviews, our 33-parameter classifier is in the range of published systems that also use speech alone, including some far larger ones.

The score shown is F1 for the depressed group: it balances how many people with depression symptoms are found against how many of the people flagged really have them. Higher is better.

We compare only with speech-only systems. Systems that read transcripts need a written record of every word, and can pick up clues from the interviewer's questions.

F1 for the depressed group, DAIC-WOZ, speech only
  • Valstar et al. 2016 challenge baseline 0.41 test set
  • Ma et al. 2016 0.52 dev set
  • Al Hanai et al. 2018 0.63 dev set
  • Wu et al. 2023 large speech models 0.70–0.80 dev set
  • Ours 33-parameter classifier 0.53–0.65 cross-validated · test set

Dev set: the development data used while building a model, which usually scores higher than an unseen test set. Test set: held-out data. Ours: 0.53 cross-validated, 0.65 on the held-out test set.

Open source

pyvoicebox

The pitch and voice-steadiness measures are computed with pyvoicebox, our Python port of VOICEBOX, the speech-processing toolbox by Mike Brookes at Imperial College London. It is open source (LGPL-3.0) and free to use.

Next steps

A Clinical Pilot in Malay

The device listens to the sound of speech, never to the words, so it is designed to work across languages. The next step is to test and validate it with Malay speakers, in a clinical pilot at IIUM Kuantan.

Listen early. Refer sooner.

Team

International Islamic University Malaysia

  • Muhammad Fahreza AlghifariInformation Systems, Kulliyyah of Information and Communication Technology[email protected]
  • Teddy Surya GunawanElectrical and Computer Engineering, Kulliyyah of Engineering[email protected]
  • Mira KartiwiInformation Systems, Kulliyyah of Information and Communication Technology[email protected]
  • Ali SophianMechatronics Engineering, Kulliyyah of Engineering[email protected]
  • Jamalludin Ab RahmanCommunity Medicine, Kulliyyah of Medicine[email protected]

Funded by the Prototype Development Research Grant Scheme (PRGS25-029-0073), Ministry of Higher Education Malaysia.