Malayalam voice assistant, on CPU only

Ask a question in Malayalam. Three small models answer it on this Space's CPU, with no GPU and no external API:

1. Speech to text: my fine-tuned Whisper-small, run with whisper.cpp (ggml-model-q8_0.bin)
2. Reply: gemma-3n-E2B-it-Q4_K_M.gguf, a small open model run with llama.cpp
3. Text to speech: my fine-tuned MMS-VITS voice

The timing table shows where the time goes. The reply model is the weakest link: small open models write limited Malayalam.