macOS Big Sur issue: Cepstral Voices not recognised
Posted: Tue Aug 09, 2022 9:04 pmHello all,
I see a certain lack of recent posts here on the forums, so I'm not counting on a quick reply, but... it's always worth a try.
Eons ago, when I upgraded macOS to Catalina, all Cepstral voices stopped working. I opened a ticket, and I actually got an answer, saying that an upgrade to the 64bit-only version of macOS was being actively worked on and would be released soon.
In the meantime, Apple's own voices had been dramatically improved, so I forgot all about the valid Millie license I still had paid for. Today, by chance, I was updating old forum passwords (including this one!) and remembered to attempt to install Millie again.
It worked flawlessly; the installer seems to be properly signed and notarised by Apple (not so common these days with 'old' software!) and, as expected, the familiar Preference Pane for Cepstral Voices was correctly installed. I could easily add the license and unlock the Millie voice, just as before the macOS Catalina days, and everything seems to be operational.
Tough luck: there is nothing on the system that recognises the existence of Cepstral Voices — except that Preference Pane. In other words, Apple must have moved on with their own voice integration scheme, and whatever scheme Cepstral used to have, it lost compatibility with contemporary versions of macOS. Bummer. On top of that, the Preference Pane doesn't even recognise Dark Mode (which is actually surprising!). From my little understanding of native Mac development, I can only guess that there are new layers of software that didn't exist when Cepstral was releasing its first versions, and access to the system voice commands is only possible when linking with contemporary libraries. There is sadly nothing I can do about that — unless I've missed to click on a checkbox somewhere. That's certainly a possibility and the reason I thought to ask in the forums first before submitting a ticket.
swift seems to have been properly installed as well. It is a universal binary — popular again, because of the new Apple Silicon chips — but it only provides support for two architectures, i386 (32 bits — there is no contemporary version of macOS still running that) and x86_64 (64 bits). There is no ARM support yet (required for Apple Silicon chips!), but since I've got an incredibly old PowerBook Pro, I'm good. Elsewhere in these forums, I've described how I managed to get my Raspberry Pi to successfully run the ARM swift version (but I have no valid license for that). However, all the new Macs use 64bit ARM CPUs, and it's unlikely they'll be able to run 32bit code, which is the only code that swift supports under ARM.
In any case, as said, that doesn't affect me (yet!), it's just the lack of integration with the system voices that annoys me. Therefore, I played a bit with swift, just to make sure that a) the Millie voice was operational; and b) I could actually do the same from the command line as I was used to do with a system-installed voice. To my surprise, the answer seems to be 'no': allegedly, the license for macOS only covers the GUI version and not the command-line version (i.e. swift)!
This is a bit baffling. I mean, what's the point of having the macOS installer also installing swift, if it cannot use the same licenses as the GUI?
Since there is no specific license for 'CLI version of macOS', I wonder what would be the appropriate license to buy? Possibly a Linux version? If that's the case, perhaps this could be made a bit clearer on the website, by labeling the version as specifically as 'macOS CLI'. It seems unfair to charge for two licenses for personal use, if one wishes to run text-to-speech experiments to demonstrate Cepstral's voice engine and its capabilities... I'd have thought that licenses were charged per seat, not per different application used. (i.e. GUI vs. CLI — if you wish to use both, that means buying two licenses. Am I right?)
That said, when testing the Millie voice — running in 'demo' version — it sounded far worse than what I remembered. Granted, it's possible that I'm not used to listening to the demo version. But the quality was decidedly far inferior to even the 'low-quality' versions of the Apple system voices: there seem to be so few samples per second that the voice sounds 'broken', clearly a recorded voice — and a badly recorded one, at that — nothing like what I remembered. I tried to record a sample on my Raspberry Pi as well — the sound quality (also unlicensed, i.e. demo version) was as bad as the one from the Mac. In fact, it was so bad, that I got far better results with the command-line version of the Mac's Siri voice (the built-in say command). The Cepstral Millie voice is so much better at intonation and expression, but its quality... argh. I hope it's just because it's a demo — since, these days, you can use the Google TTS service, which also uses SSML, and has an uncanny quality, with 220 voices for 40 different languages, all of that for a very reasonable price (the first million characters per month are for free)...
I see a certain lack of recent posts here on the forums, so I'm not counting on a quick reply, but... it's always worth a try.
Eons ago, when I upgraded macOS to Catalina, all Cepstral voices stopped working. I opened a ticket, and I actually got an answer, saying that an upgrade to the 64bit-only version of macOS was being actively worked on and would be released soon.
In the meantime, Apple's own voices had been dramatically improved, so I forgot all about the valid Millie license I still had paid for. Today, by chance, I was updating old forum passwords (including this one!) and remembered to attempt to install Millie again.
It worked flawlessly; the installer seems to be properly signed and notarised by Apple (not so common these days with 'old' software!) and, as expected, the familiar Preference Pane for Cepstral Voices was correctly installed. I could easily add the license and unlock the Millie voice, just as before the macOS Catalina days, and everything seems to be operational.
Tough luck: there is nothing on the system that recognises the existence of Cepstral Voices — except that Preference Pane. In other words, Apple must have moved on with their own voice integration scheme, and whatever scheme Cepstral used to have, it lost compatibility with contemporary versions of macOS. Bummer. On top of that, the Preference Pane doesn't even recognise Dark Mode (which is actually surprising!). From my little understanding of native Mac development, I can only guess that there are new layers of software that didn't exist when Cepstral was releasing its first versions, and access to the system voice commands is only possible when linking with contemporary libraries. There is sadly nothing I can do about that — unless I've missed to click on a checkbox somewhere. That's certainly a possibility and the reason I thought to ask in the forums first before submitting a ticket.
swift seems to have been properly installed as well. It is a universal binary — popular again, because of the new Apple Silicon chips — but it only provides support for two architectures, i386 (32 bits — there is no contemporary version of macOS still running that) and x86_64 (64 bits). There is no ARM support yet (required for Apple Silicon chips!), but since I've got an incredibly old PowerBook Pro, I'm good. Elsewhere in these forums, I've described how I managed to get my Raspberry Pi to successfully run the ARM swift version (but I have no valid license for that). However, all the new Macs use 64bit ARM CPUs, and it's unlikely they'll be able to run 32bit code, which is the only code that swift supports under ARM.
In any case, as said, that doesn't affect me (yet!), it's just the lack of integration with the system voices that annoys me. Therefore, I played a bit with swift, just to make sure that a) the Millie voice was operational; and b) I could actually do the same from the command line as I was used to do with a system-installed voice. To my surprise, the answer seems to be 'no': allegedly, the license for macOS only covers the GUI version and not the command-line version (i.e. swift)!
This is a bit baffling. I mean, what's the point of having the macOS installer also installing swift, if it cannot use the same licenses as the GUI?
Since there is no specific license for 'CLI version of macOS', I wonder what would be the appropriate license to buy? Possibly a Linux version? If that's the case, perhaps this could be made a bit clearer on the website, by labeling the version as specifically as 'macOS CLI'. It seems unfair to charge for two licenses for personal use, if one wishes to run text-to-speech experiments to demonstrate Cepstral's voice engine and its capabilities... I'd have thought that licenses were charged per seat, not per different application used. (i.e. GUI vs. CLI — if you wish to use both, that means buying two licenses. Am I right?)
That said, when testing the Millie voice — running in 'demo' version — it sounded far worse than what I remembered. Granted, it's possible that I'm not used to listening to the demo version. But the quality was decidedly far inferior to even the 'low-quality' versions of the Apple system voices: there seem to be so few samples per second that the voice sounds 'broken', clearly a recorded voice — and a badly recorded one, at that — nothing like what I remembered. I tried to record a sample on my Raspberry Pi as well — the sound quality (also unlicensed, i.e. demo version) was as bad as the one from the Mac. In fact, it was so bad, that I got far better results with the command-line version of the Mac's Siri voice (the built-in say command). The Cepstral Millie voice is so much better at intonation and expression, but its quality... argh. I hope it's just because it's a demo — since, these days, you can use the Google TTS service, which also uses SSML, and has an uncanny quality, with 220 voices for 40 different languages, all of that for a very reasonable price (the first million characters per month are for free)...