Yup, the fallback can’t be sped up particularly. It’s actually vastly faster since 1.1 and just needs a fraction of a second now, but as you’ve observed it is a (very) stripped-down form of speech synthesis so there’s no getting around the CPU load and time.
If this were my thing I think I would keep my own Core Data model of the songs on the device along with their already-created pronunciation, and on launching the app I’d do a comparison of the user’s songs and my database of the user’s songs and just add pronunciations for new ones.