We use nice TTS (imo) for the dictionary and example sentences, for ‘text mode’ we fall back to browser TTS which is… not nice in Chrome, much better in MS Edge browser. We experimented with making our own voices, actually recorded a dataset last year for English to release that was better than LJSpeech. We’ll try make improvements.
Pause behaviour needs some attention, you are correct.
RE: ‘streamlined exports that could tie new words to the source’, how are you exporting words? With HTML export, or to Anki?