use "utf-8" encoding to fix TTS inference

[eschmidbauer]

This enables inference from model_inputs/resynthesis_prompts.json to be non-phonemized text.
example

    {
      "script": "While complexity makes you sound smart, focusing on the key basics matters more.",
      "spk_id": "ljs",
      "decoder_spk_id": "ljs",
      "duration_spk_id": "ljs",
      "energy_spk_id": "ljs",
      "f0_spk_id": "ljs",
      "language": "en_US",
      "emotion": "other"
    }