For our new mobile language-learning app I need a production-ready machine-learning model that listens to a learner’s microphone input and instantly returns two things: (1) numeric accuracy scores at both word and whole-sentence level and (2) a simple “good / poor” quality label. The first release targets English for students in roughly Grades 6-9, and it must handle both isolated words and connected speech so the app can coach through single-word drills and free-speaking exercises alike. You’ll own the full ML pipeline—from choosing an acoustic/phoneme-alignment backbone (e.g., wav2vec, HuBERT, DeepSpeech, Whisper, or any framework you trust in PyTorch, TensorFlow, or ONNX) to training, evaluation, and exporting an inference-ready artifact. Latency and scalability matter because thousands of real users will hit the API, so batching, streaming, or lightweight on-device options should be considered. Deliverables • A trained pronunciation-scoring model with documented hyper-parameters and training script • Inference code that exposes a clean API (JSON in/out) returning word-level scores, sentence-level score, and the categorical label • A brief report outlining data used, evaluation methodology, accuracy metrics, and recommendations for threshold tuning • Instructions for dockerised deployment so our engineering team can plug it straight into the existing backend Once we prove the English launch, we plan to extend to Spanish or Mandarin and potentially move toward a longer-term collaboration. After you review the general scope, let’s jump on a quick call to align on dataset access, target correlation with human raters, delivery timeline, and milestones.