Indian Speech-to-Text Model Customization

I need open-source speech recognition stacks fine-tuning for high accuracy, reliable, fully on-prem deployment across seven Indian languages—Hindi, Tamil, Bengali, Marathi, Telugu, Kannada and Gujarati.
My current shortlist of engines is Whisper, AI4Bharat’s Indic models and Vanni from IISc; you are free to combine or layer these as long as the final pipeline remains 100 % open source and runs locally on our GPU servers (Windows CUDA, PyTorch).

What I want from you is a measurable lift in word-error-rate, inference speed and resource efficiency. Techniques such as transfer-learning, domain-specific vocab augmentation, quantisation, batching or ONNX/TensorRT optimisation are all on the table—choose what works best for each language pair.

Deliverables
• A reproducible training & fine-tuning script set (shell + Python notebooks)
• Containerised inference service with REST/gRPC endpoints, ready for on-prem deployment
• Benchmark report comparing pre- and post-tuning accuracy, latency and memory footprint for every target language
• Brief “getting started” document so my internal team can extend or retrain the models later

Acceptance criteria
• ≤ 10 % relative WER reduction over the base model on our internal test set for at least five of the seven languages
• End-to-end transcription latency under two seconds for a 15-second audio clip on an RTX A4000
• All code and configs licensed under compatible open-source terms

If you have prior experience optimising Whisper or Indic ASR models and can point me to benchmarks you’ve achieved, I’m ready to review proposals right away.