Sorry to post into this thread so late but I am getting to the point where we as a community need to demand that this get better, and discuss clear reasons why this is happening.
I have a clear picture of how voice dictation worked when the iphone 4s came out, and what has happened since. I am fairly convinced a variety of factors came into play to destroy our user experience.
1) Apple consumed the Siri team, and a cultural change happened.
2) where it didn't want to happen, people simply quit, leaving with them their knowledge
3) somebody at apple started driving a different more ML (machine learning) type of way.
4) repeat 2)
5) Feature requests came in and the #1 feature request would be 98% of the feature requests: Respond to my
-name
-town
-country
-favorite movie
And a million other "proper nouns". The result is proper nouns that are never spoken by you getting substituted for real common phrases ("Wei" being substituted for "way"... "Dash" being substituted for "-"... etc.
IMHO This one is what broke things the most.
6) A context engine was added, but it is not good. The context engine fixed some bugs while causing many others. It doesn't take into account all the context available(a very simple example is "how many syllables did I receive"). Other missing context. Another context missing is related to sound-alike word meanings and resolving those. Words like "except" always substitute for "accept". "to" always substitutes for "too" (IBM voice rec could do this correctly in the early 90's). These types of context need to follow programmed rules, and should NOT rely 100% on machine learning. Machine learning should be used to understand "sounds"; the rules of English should be BAKED.
7) The team did not maintain a separate and elaborate data set (many spoken voices) to perform recognition on, which developers should not have access to. The developers probably have a regression testing system that everyone has access to. So they fix all the bugs in that set, while introducing other flaws. This is an absolute MUST with machine learning- there has to be regression data that no one has access to or else you will hard code to that data and create code that is not general enough to handle the larger data set of "everyone in the world". A separate example: nVidia changing performance characteristics of their rendering code when they sense the "quake test" is being run. Or volkswagen tuning their diesel engines to run clean when idling (which is the case during a DMV test). So all the test data looks great, but real world usage is terrible.
8) I think at some point because of the much larger world of english speakers and thus much lower quality of english speaking, Apple decided that their context engine should not just work hard to get what the speaker said... but also to correct grammar. For this reason, I find the context engine "goes wild" largely altering my thoughts completely. Because new english speakers may literally utter "the wrong words", Siri is now allowed to correct the results to "what you should have said". for this reason, I am seeing positives turned to negatives and complete phrases added to the output that have nothing to do with / sound nothing like the words I just uttered. IMHO this was a huge mistake.
8) The team failed. It was unable to solve the problems. Some people gave up in frustration and left after banging their heads against the wall, while others stuck in but gave up on their sense of urgency to make things right. This is what living in Apple QA is all about. Less respect for the testers and the "specification" is whatever the developers say is OK. When you have one or two bad apples on the developer team (yuk yuk) everything falls to heck.
9) Since Siri voice dict, while the worst product in Apple's portfolio, is a small piece of the larger picture, it is not enough to sink the ship. This was declared good enough, and the only bug fixes that are going in are "hey our new (proper noun) isn't getting recognized- fix it it's high priority and urgent." This problem is not specific to Apple, nor are most of these issues.
10) And finally, We are now seeing in the software world some of the same behavior that destroyed the oil industry, the banking industry, the automobile industry, and basically everything in the past: we have a monopoly. Apple / Google are the only competitors left in this space, and each has their strengths. As a result, there is very little chance that small problems will get fixed. This is on US as consumers, not to demand better software and hold these companies accountable.
The net result is that the recognition rate for many common sentences seems to have gone below 50%.
As an aside, while the siri assistant is still pretty good (because it uses a very strong context/limited choices) of "available commands", it can still be affected by once awesome now terrible recognition engine. Therefore using siri while driving no doubt is resulting in accidents and other tragedies (the lesser ones being people breaking their iphones by throwing them).
If there was an alternative for IOS, I'd be using it. I will use any new system that comes in that is better... in a heartbeat.
