On-Device Emotion Recognition for Indian Languages

A small, mobile-friendly emotion recognition model that runs inside our app and classifies the speaker's emotion across Indian languages into one of four classes: • Neutral • Happy • Sad • Angry Hard requirements: very small model size, low latency, and works across multiple Indian languages (Hindi, Tamil, Telugu, Bengali, Marathi, Kannada, Malayalam, Punjabi, Gujarati, etc.). Size & performance targets • Model size: ideally under 10 MB on disk after quantization (smaller is better) • Latency: inference under ~100 ms per sample on mid-tier mobile • Runs fully on-device — no server round-trip • Accuracy: meaningfully better than chance across all target languages, not just one Where it runs:: • iOS (CoreML) and Android (TFLite / ONNX Runtime Mobile / NCNN) • Quantized (INT8 or FP16) for size and speed Approach — open to your judgment We're not prescribing the architecture. We expect you'll evaluate options like a distilled wav2vec2 / HuBERT / WavLM head, a lightweight CNN or CRNN on mel-spectrograms, or a custom small model. Use whatever public Indian-language SER datasets you can (IIT-KGP SEHSC, IITM, IndicSUPERB, Hindi/Bengali/Tamil emotion corpora, etc.) plus augmentation. Tell us your plan. You should apply if:: • You've built speech / audio ML models before — ideally speech emotion recognition or paralinguistics • You've shipped a model to a mobile device, not just a notebook • You understand model compression: quantization, pruning, distillation • Bonus: you've worked with Indian-language speech data and know its quirks (code-mixing, accents, low-resource constraints) Deliverables:: • Trained model + training/eval code • Mobile-ready model files (CoreML + TFLite/ONNX) • Inference snippet / module we can drop into our app • Benchmark report: size, latency, per-language accuracy, confusion matrix • Short README on how to retrain / extend to new languages Please include in your proposal • A 2–3 sentence pitch on your approach (base model, datasets, compression plan) • Which Indian languages you're confident you can cover well • Links to past work (repos, demos, papers, apps) • Realistic timeline and budget for a working prototype