AI Engineer / Research Engineer for Multilingual Indian TTS Model

We are looking for an experienced AI/ML engineer or researcher to build a suite of lightweight Automatic Speech Recognition (ASR) / Speech-to-Text (STT) models focused on Indian languages. The primary objective is to develop individual language-specific STT models, with each model targeting a size of under 100 MB, while maintaining good transcription accuracy, low latency, and efficient CPU inference. The models should be suitable for deployment on edge devices, Linux servers, and future Android applications. The project should include: * Development of production-ready STT models for individual Indian languages. * Each language should have its own independent model (rather than a single multilingual model). * Support for languages such as Hindi, Telugu, Tamil, Kannada, Malayalam, Bengali, Marathi, Gujarati, Punjabi, Odia, Assamese, and additional Indian languages where feasible. * Support for real-world speech, including different accents, speaking speeds, and background conditions wherever supported by the available datasets. * Accurate handling of punctuation, numbers, dates, currency, abbreviations, and commonly spoken English words mixed into Indian language conversations. * Low-latency streaming inference is preferred. If streaming and offline inference require different architectures, please describe the recommended approach. * Training using open-source datasets wherever possible. Commercial datasets may also be recommended if they provide a significant improvement in recognition accuracy, accent coverage, robustness, or language coverage. Any recommendations should include licensing information, estimated costs, and technical justification. * Complete reproducible training pipeline, including dataset download, preprocessing, audio normalization, text normalization, tokenizer preparation, training scripts, evaluation, checkpointing, and model export. * A complete fine-tuning pipeline that enables adding new datasets, adapting to new domains, improving recognition accuracy, continuing training from checkpoints, and retraining individual language models. * Comprehensive documentation covering environment setup, dependencies, dataset preparation, training, fine-tuning, evaluation, ONNX export (preferred), streaming inference, offline inference, model architecture, hyperparameters, hardware requirements, deployment, and troubleshooting. * Delivery of all source code, configuration files, dataset download scripts, preprocessing scripts, training scripts, fine-tuning scripts, trained model weights, exported inference models, example inference applications, and complete documentation. Candidates should have strong experience with modern ASR architectures such as Whisper, Parakeet, Conformer, Citrinet, NeMo, wav2vec 2.0, MMS, or other state-of-the-art speech recognition models. Experience with PyTorch, ONNX, Hugging Face, NVIDIA NeMo, multilingual speech processing, and Indian language ASR is highly preferred. Please include in your proposal: * Similar STT/ASR projects you've completed. * Languages you have previously worked with. * The model architecture you recommend and why. * Expected model size for each language. * Expected Word Error Rate (WER) or Character Error Rate (CER) for each language, if available. * Whether streaming inference is supported. * Hardware requirements for training. * Estimated timeline. * Estimated budget. * Public GitHub repositories, research papers, demos, or previous work. The final solution should be production-ready, maintainable, and optimized for deployment. Preference will be given to candidates who have demonstrable experience building lightweight, high-accuracy speech recognition models for Indian languages and can justify architectural decisions with respect to model size, latency, accuracy, and deployment efficiency.