This is the start of the #EloquenceResearchCommentary thread. This one will be primarily used for my own personal thoughts on the synth, as well as smaller or more technical findings that might not quite fit into the main research findings thread. These will probably be shorter posts with a much less formal tone.
Funny how IBM added the pronunciation of Enron to IBMTTS around February 2003, about a year or so after the company collapsed in spectacular fashion. They also fixed Aptiva, the name of a line of computers they produced, the pronunciation of which had been broken since 2000, about a year after they discontinued them. #EloquenceResearchCommentary
So Eloquence's maximum pitch is 422 Hz? The Korean version sincerely begs to differ... You might want to turn your volume down or wear headphones for this. This is the result of setting the pitch and inflection settings to their maximum, holding down the i key for several seconds, and then adding an exclamation mark. The Korean version of IBMTTS, fortunately or unfortunately, depending on how you look at it, fixed or didn't get this, but still screams a bit. #EloquenceResearchCommentary
An 1996 article from The New York Times about Everybody can Read, a program by Lew Robins that was designed to teach children to read and used Eloquence. Also contains some early audio samples of the synth. https://archive.nytimes.com/www.nytimes.com/library/cyber/week/0926read.html
#EloquenceResearchCommentary
The Infinitely Patient Computer Teaches Problem Students to Read

On some products that used Eloquence, such as the BrailleNote from HumanWare or the PAC Mate from Freedom Scientific, there's a strange arrangement of libraries. The US English library, enu.syn, is running Eloquence 6.0, while all other languages are at 6.1. This isn't exclusive to just these products. I've seen a similar thing with COBRA, an old German screen reader. All languages are at version 6.1, except for US English which is at 5.0.1.8. I wonder if this was done for performance reasons or to save on licensing costs. #EloquenceResearchCommentary
RealPlayer. A media player written in the 90s, significant enough to be added to the 3,518 words the Japanese Eloquence won't spell out, as well as the German IBMTTS in version 6.4.1.0. Neither Eloquence or IBMTTS added it to English, though, so the letter to sound rules assume that the root word is alplay, with "re" as the prefix, and "er" as the suffix. There's some support for mixed case words, example: LargeBoulder, but even that doesn't catch it. #EloquenceResearchCommentary
Prior to around March 1997, ETI didn't have a website. However, the domain existed at least as far back as December 1996, and was briefly the home of Ithaca Guitar Works. https://web.archive.org/web/19961227210037/eloq.com/. The Ithaca Guitar Works site is well-archived, even including audio and video samples of some of the instruments they were selling at the time. Here's a sample of the mandolect. https://web.archive.org/web/19970203141736/http://eloq.com/wav/mando.wav. The last capture where the domain was hosting Ithaca Guitar Works was February 1997. https://web.archive.org/web/19970203141717/http://eloq.com:80/. The next capture was made in April 1997, and the domain had switched to ETI at that point. https://web.archive.org/web/19970412190013/http://eloq.com:80/. Since Ithaca Guitar Works and ETI were in the same city, I'm wondering if ETI gave Ithaca Guitar Works permission to use their domain while they were setting up their own site.
Ithaca Guitar Works still exists today. https://www.guitarworks.com/. #EloquenceResearchCommentary
Ithaca Guitar Works

A random crash string that worked in the British English version of
Eloquence 4: ironasap. This one is interesting. ASAP itself was added to
Eloquence in 1999. For US English, it's pronounced as a-sap, but for
British English, it says each letter individually. However, it does this
using an entry in the internal roots dictionary which has multiple stress
markers. The SPR is `[.2e.2E.2se.1pi]. It seems that the bug is that if a dictionary entry contains an SPR that has multiple stress marks, and certain known word prefixes are used, the synth crashes. For ASAP, apart from iron, other known prefixes I've found to work are cyber and body. I wanted to see if this worked in the user roots dictionary by adding the SPR for IBM's powerpc dictionary entry, which has multiple primary stress marks. Ironpowerpc didn't crash, but the iron prefix got stripped out, though the SPR for powerpc was still modified to fit with the addition of the iron prefix, which means Power PC ended up sounding more like power pissy instead. This bug was fixed in Eloquence 5, so typing ironasap in the British English version of Eloquence today just results in a weird stress pattern, with the ern part of iron getting primary stress. #EloquenceResearchCommentary
@MutedTrampet Dam that's hella interesting, didn't know this synth worked like that.
@TheQuinbox @MutedTrampet Perhaps there was some phoneme or phonemes considered better-sounding in the old version.
@MutedTrampet What does Cobra sound like?

@MutedTrampet It's interesting that the verson of eloquence we are all used to is quite ... fixed?

What I mean is, there are loads of DECtalk flavours floating about, yet I don't recall Eloquence changing much since the big move from 4.7. My ECI.dll that I am running today has copyright 1995-2002.

And I guess part of the longevity it enjos is it's perminence thorugh all the other changes? I dunno. Just rambling now.

@cachondo Yeah, the stability of the voice could be a part of what keeps it going. The PC version of Eloquence has pretty much never been recompiled since July 30, 2002, and I'm not aware of any major voice developments from SpeechWorks or Nuance after the introduction of the 16k sampling rate in the Symbian version. Maybe SpeechWorks decided it wasn't worth doing much with after that point? Who knows? To be honest, though, I wouldn't be surprised if not much can be done to develop the synth further at this stage, even if someone wanted to. The 90s papers mention that Eloquence was written in its own Delta programming language and compiled to C, so the source code might very well be a bunch of unreadable data tables that would require the original Delta files to understand.
@MutedTrampet @cachondo Considering that Apple is making its changes via dictionaries and not via modifying the original synths, I suspect that either the Delta code / tools have been lost, or are very deliberately not given out to customers.
@miki @MutedTrampet @cachondo Maybe I'm not understanding something here, if source code is not available, how possibly they could compile it to work in 64 architecture on IPhones??

@kaveinthran @MutedTrampet @cachondo As far as we know, there are two "stages" to the Eloquence compilation process.

Stage 1 compiles the source code written in Delta, a language to nicely express rules for transforming text into phonemes, into C++.

Stage2 takes that C++ code for stage1, along with the rest of Eloquence, which is also written in C++, and compiles that to whatever architecture it's supposed to be running on, just like a normal C++ program.

Apple can definitely do stage2, but we don't know whether they can do stage 1.

@MutedTrampet what's the difference between say, eloq 5x, 6 6.1?
@kaveinthran Eloquence 5 is pretty much the voice we know today, but with a somewhat different intonation pattern and less pronunciation fixes. Eloquence 6.0, which I oddly can't find anywhere except notetakers, has Eloquence 5 intonation, but many new dictionary entries and letter-to-sound rule changes. Eloquence 6.1 was a touch-up of sorts for the work done in Eloquence 6.0, fixing the pronunciation of many words and names, example: liam, and fixing some long-standing bugs, such as the phrase "the concrete floor" choosing the wrong form of concrete.
@MutedTrampet Wow, thanks, Which devices or screenreaders that you know of is using eloquence 6? And what sort of letter-to-sound rule changes is done in v6.0?
@kaveinthran The devices I know of that use Eloquence 6.0 are the PAC Mate from Freedom Scientific and the BrailleNote from HumanWare. Many changes were made to the letter-to-sound-rules in 6.0. One of the changes I know of is that the string ane, if it occurs at the end of a word, will be pronounced like ain rather than an, which ended up fixing words like halothane. Another change that was made was that if a word ends in ante or ente and it has at least one consonant preceding those two strings, ante and ente will be pronounced ontay and entay, respectively.
@MutedTrampet @kaveinthran this makes me wonder what version jaws uses?
@EmeraldRose @kaveinthran It uses Eloquence 6.1, created on July 30, 2002.
@MutedTrampet @kaveinthran Ah ha, I didn't know that. Nice. And this may be the version we all know and love, which still doesn't pronounce some words correctly.
@EmeraldRose Yep, 6.1 is the version pretty much everyone has. It was the last version of Eloquence ever made for PCs.
@MutedTrampet Wow, i know in jaws they have fixed pronunciation errors but with other synthes, they haven't.
@MutedTrampet wow, thanks for poasting this, as a tts nird, I have always wondered how old eloquence including the eddy voice sounded like, and now I know
@MutedTrampet I thought that Eddie was always more breathy. The audio sample of the first paragraph says it's Eddie but it sounds more like Wade to me.
@BTyson Could be an early representation. The oldest builds we have on PCs are 3.x builds from 1997, but older versions existed.