Selected work
Walaw AyaAugust 2025 – May 2026

Recitation practice.On-device feedback.

Quranic Recitation & Tajweed Error Detection App

An Android prototype for Quranic recitation practice, connecting a Flutter interface with local speech recognition and word-level feedback.

My contribution

Mobile development, native integration, local inference, audio preprocessing, and system testing within a senior design team.

Speech → understanding
Audio inputOn-device inference

Conceptual project illustration

Character error rate
0.294%
Word error rate
1.389%
Tajweed F1
≈98.7%

Team-level validation results. CER and WER refer to base transcription; these are not production or real-world user benchmarks.

The problem

Practicing Quranic recitation involves both recognizing the words and attending to Tajweed. Our senior design project explored how a mobile application could give learners feedback on recitation and Tajweed errors, with speech inference running on the device.

System overview

  1. 01Audio inputRecitation
  2. 02Audio processingPreprocessing / Silero VAD
  3. 03Local inferencewhisper.cpp / Dart FFI
  4. 04Recitation feedbackComparison / Flutter UI
High-level system overview, based on the project description. Inference runs locally on the Android device.

A complete practice workflow

The Android prototype supported verse-by-verse and continuous recitation. Word-level mismatch highlighting and an error-review flow helped users revisit their recitation, while reference audio provided a point of comparison.

  • Verse-by-verse and continuous recitation
  • Word-level mismatch highlighting and error review
  • Reference audio playback and configurable settings

Connecting Flutter to local inference

I worked on the software integration between the Flutter/Dart interface and native whisper.cpp through Dart FFI. The system also incorporated Silero VAD and local audio preprocessing.

The on-device approach removed the need for a remote inference server. It also made the boundaries between mobile UI, audio processing, and native inference central to the implementation and testing work.

Models and validation

The project used Whisper-based Quranic ASR models fine-tuned on recitation data from 25 professional reciters. The team’s validation results were 0.294% character error rate and 1.389% word error rate for base transcription, with approximately 98.7% Tajweed F1.

An AUS Undergraduate Research Grant of AED 5,000 supported the project in 2026.

Scope and limitations

Walaw Aya was a team-built senior design Android prototype. The reported metrics are from the validation set and should not be read as evidence of performance across every speaker, recording environment, or Android device.

The work brought together mobile development, native integration, speech processing, and testing in a single application.

Technologies

  • Python
  • Flutter
  • Dart
  • whisper.cpp
  • Dart FFI
  • Silero VAD
  • ASR
  • Deep Learning
Next case studyAmazon recommendations
Open to graduate & entry-level opportunitiesView résumé

Let’s talk
about what’s next.

If you’d like to talk about software, AI, or an opportunity, I’d be glad to hear from you.