At length_scale 1.0 the models speak too fast. This sets model_args.length_scale in each config.json:
- Male (Tamil and Sourashtra script): 1.43 (target 0.7x of the current speed)
- Female (Tamil and Sourashtra script): 1.11 (target 0.9x of the current speed)
inference.py reads the value from config.json, so no code change is needed. Only this one value changes in each config. The README gets a short Speaking Speed section explaining it.
