Crystal Clear Memories
Computer Vision
3D GenAI
Automated transformation of photos into 3D memorial keepsakes using a hybrid of classical computer vision and 3D generative AI.
Cases
NDA
AI-Powered Pronunciation Correction for Children’s Speech Therapy
Key results
86%Achieved 86% accuracy in phoneme recognition
- Reduced the manual workload for speech therapists
Services- Audio processingSpeech recognition
Industry- Speech therapyHealthcare
TECHNOLOGIES USED- PythonAWSTerraformFastAPIGradioTransformersPyTorchAccelerateWav2Vec2
Team- 2-4 AI engineers
Duration- 1 year
Location- United Kingdom
The Challenges & Goals
The project set out to solve a key challenge in paediatric speech therapy: automatically identifying and tracking pronunciation mistakes without constant therapist involvement. The goal was to develop an AI system that detects mispronounced phonemes in real time, handles non-linear speech patterns, and delivers feedback as accurately as – or better than – a trained therapist. But children’s speech brings a unique set of challenges:
- 01
Unpredictable articulation and speech variability
Children’s pronunciation changes as they grow, making phoneme boundaries inconsistent and difficult for traditional models to interpret.
- 02
Diverse accents and speech patterns
Regional and individual pronunciation differences required the model to handle a wide range of phonetic variations.
- 03
Noisy real-world environments
Sessions often take place outside ideal recording conditions, with background noise, interruptions, or low-quality microphones.
- 04
Limited training data for children’s speech
Few existing datasets capture the variety of children’s voices and speech errors needed to train accurate recognition models.
Solution Overview
We brought the vision to life through a multi-stage approach, from research and model selection to dataset creation and deployment. Each step addressed specific technical and data challenges while aligning with therapy needs. Using advanced speech recognition, phoneme-level training, and a custom child speech dataset, we built an AI system that delivers reliable, actionable feedback even in noisy, real-world environments.
- Stage 01
Framing the Problem and Model Selection
The client’s vision was clear: automate pronunciation correction for children. But traditional speech recognition models weren’t designed for the nuances of kids’ speech, especially with accents, articulation variability, and noise.
What we did:
- Identified phoneme recognition as the most effective approach for tracking mispronunciations.
- Selected Wav2Vec 2.0, a self-supervised model known for handling raw audio and noisy data.
- Proposed a fine-tuning strategy tailored to children’s speech patterns.
This gave the project a clear path forward and the confidence that the goal was technically achievable.
- Stage 02
Adapting Wav2Vec 2.0 for Children
Wav2Vec 2.0 had strong potential, but it needed careful fine-tuning. Children’s speech is less structured and more variable than adult audio, which makes direct transcription difficult.
Key steps:
- Fine-tuned the model using Connectionist Temporal Classification (CTC) loss, aligning unsegmented audio with phoneme sequences.
- Integrated a multilingual variant of Wav2Vec 2.0 to handle diverse phonetic inputs.
- Added logic to handle non-linear, irregular speech patterns, making the model more robust.
The result was a fine-tuned model capable of understanding real-world, imperfect children speech.
- Stage 03
Building a Custom Dataset
Data availability was a critical roadblock. Existing datasets lacked both the quantity and diversity needed for children’s phoneme-level speech.
How we addressed it:
- Built a custom dataset with recorded child speech across varied conditions.
- Partnered with speech specialists for accurate phoneme-level annotation.
- Created a validation pipeline to automate checks and reduce expert workload.
- Applied feature engineering to capture markers of disordered or unclear speech.
The dataset became the foundation that enabled real-world accuracy rather than just lab success.
- Stage 04
Delivering the Solution
With the model trained and validated, we moved into deployment, focusing on usability, monitoring, and production-readiness.
Key results:
- Achieved 86% accuracy, outperforming the 70% baseline of manual therapist assessment.
- Delivered actionable phoneme-level feedback for children’s pronunciation.
- Created a production-ready inference pipeline with built-in analytics.
- Deployed a CI/CD system to streamline future updates and improvements.
This collaboration demonstrated how AI and computer vision can bring lab-level insights into users’ pockets, redefining how personal health can be tracked and understood.
Let’s Talk About Building AI That Works in the Real World
No vague promises. No pressure. Just a real conversation with our technical leadership to understand your needs and explore the best next steps.
What Happens Next?
- Your message goes straight to our core team – not a sales department.
- You’ll talk directly with our CEO, CTO, or a domain expert for a consultation-style first call.
- We’ll ask the right technical questions to understand your challenge and scope.
- If we’re a fit, we’ll follow up with a proposed path forward.
More Projects Like This
Pardigm
Computer Vision
Edge AI
Hormone tracking application powered by Computer Vision with 93%+ accuracy in real-life conditions.