Urdu Speech Therapy Practice Game for Children with Therapist Review and Sound Scoring

The problem: therapy happens once a week, practice should happen every day

Children with speech sound disorders, such as difficulty producing certain consonants, improve with frequent, targeted practice. A therapist usually sees a child for a short session once or twice a week, then asks parents to practise at home every day. In practice, home practice is irregular and boring for the child, and parents are unsure whether the child is saying the sound correctly.

In Pakistan the problem is sharper. Qualified speech and language therapists are few and concentrated in large cities, so families may travel far for a session and get little support between visits. Most speech therapy apps are built for English, with English word lists and English sound targets that do not match Urdu.

What already exists and the gap

• Commercial speech apps for children such as Speech Blubs focus on general speech development rather than therapist-set sound targets, and we could not find one built around Urdu sounds.
• Research systems such as Tabby Talks (an automated tool for assessing childhood apraxia of speech) and Apraxia World (a tablet game for child speech therapy from Texas A&M University) combine games, recordings and automatic speech analysis, but for English.
• Video calls with therapists help with access but not with daily practice.

The gap is an Urdu practice game that children enjoy, that records every attempt, gives a careful first score for a few sounds, and leaves the therapist in control of judgement and planning.

Inspiration

Tabby Talks (Parnandi et al., Speech Communication, 2015) combined a clinician interface for assigning exercises, a tablet app that prompts children and records them, and a speech recognition engine that analyses the recordings for the clinician. The later Apraxia World work (Hair et al., ACM Transactions on Accessible Computing, 2021) evaluated a tablet game for child speech therapy over a longer period. Your project adapts this design to Urdu, with a smaller automatic scoring scope and a stronger focus on the therapist's review.

What you build

• Therapists create a practice plan: target sounds, word positions (start, middle, end) and how many attempts a day. They review recordings weekly and correct scores.
• Children play a short daily game where saying a word moves a character, opens a door or feeds a pet. Each attempt is recorded.
• Parents see that practice happened and how many attempts the therapist marked as correct.

The technically interesting part

• Children's speech is hard. Speech models are trained mostly on adult voices, and children's pronunciation varies a lot. Fine-tune a small pretrained speech model on a limited set of target sounds, and report confidence instead of a confident yes or no.
• Small data, ethically collected. You will need a small labelled set of children's recordings. Work with a therapist, get informed consent from parents, store recordings securely and keep the dataset private.
• Scoring only what you can. Limit automatic scoring to three to five target sounds where your model is reliable, and leave everything else to the therapist.
• Engagement without distraction. The game must reward attempts and encourage repetition without turning practice into a race. Keep each session short, around five to ten minutes.
• On-device and offline. Run the scoring model on the phone if possible, so practice works without a connection and audio stays local until upload.

How to evaluate it

• Scoring agreement: agreement between automatic scores and therapist judgements for each target sound, reported with confidence intervals.
• Practice: number of attempts per day during the pilot compared with paper-based home practice for the same children before the pilot.
• Therapist time: minutes needed to review a week of practice through the dashboard.
• Feedback: short interviews with therapists and parents about what helped and what did not.

Suggested plan

• Weeks 1 to 4: meet one or two therapists, choose target sounds and word lists, design the game.
• Weeks 5 to 9: game, recording and upload, therapist dashboard.
• Weeks 10 to 15: consented data collection, model fine-tuning and on-device export.
• Weeks 16 to 21: pilot with at least five children over two or more weeks.
• Weeks 22 to 26: analysis, write-up and demo.

Path to a product

Therapists, special education schools and child development centres in Pakistan and other Urdu-speaking communities need tools that extend their reach. A subscription for clinics with free access for families, or a grant-funded rollout through a hospital or NGO, are realistic. The project is also a strong portfolio piece for health technology, machine learning and game roles.

Scope warning

This is a practice tool, not a diagnostic or a replacement for therapy, and the app must say so. Work only with a therapist's involvement and with parents' informed consent, and seek your university's ethics approval before collecting any children's recordings. Keep automatic scoring narrow; a therapist-reviewed game with no automatic scoring is still a valid project if the model is not reliable enough.

Learn what you need on Sefism

• Machine learning roadmap: fine-tuning and evaluating a small model.
• Game development roadmap: designing short, rewarding sessions for children.
• Speech and Language Processing: the basics of speech recognition and phonetics.
• Hugging Face Transformers Documentation: pretrained speech models and fine-tuning.
• Mobile development roadmap: recording, offline storage and running models on the phone.

Get help with this idea

The risk is a model that is not reliable enough on children's speech, or a pilot that never happens. Tauseef can help you scope the target sounds, plan the consent and data collection, and find a therapist partner. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session