Emacspeak discussion list

[aleland.tech.list AT fastmail.com]

Hey Parham,

> I’m using Emacspeak with Swiftmac and I notice that in longer lines, there 
> seems to be a queueing issue, where things get spoken out of order. I was 
> wondering if anyone has found a solution for this, or has a patch?

Yes — I hit exactly this, and it almost made emacspeak unusable for me, as 
I'm not a coder and use emacs mostly for org, writing, and notmuch, and so 
most of my emacspeak travels involved very long lines of prose. I ended up 
working through the SwiftMac source with an AI assistant (Claude, 
particularly earlier this summer when Fable was free), which did the actual 
diagnosis and wrote the fix, which I'm plagiarizing here. I was on 4.3.6 
(robertmeta/swiftmac). I feel ambivalent sending this because as a 
nontechnical vibecoder I don't fully understand what I'm saying here or even 
if this is accurate. But since I shared your problem and the solution worked, 
I will swallow my pride about ventriloquizing an LLM on a FOSS listserv and 
continue. There turned out to be two things going on:

1. The stop/interrupt path never flushes the pending text queue. SwiftMac 
queues text you send (q) and dispatches it (d) separately. When speech is 
stopped, the stop routine resets the audio player and clears the in-memory 
chunk list, but it never clears the pending command queue — so any text that 
was queued-but-not-yet-dispatched just sits there and gets spoken by the next 
dispatch, from the wrong place. (There's even a clearPendingQueue() function 
in the state store, but nothing in the code calls it — and it's 
all-or-nothing anyway, so it'd wipe your queued voice/rate/punctuation 
settings too.)

2. In-flight synthesis leak. Even after a stop, a chunk already inside 
synthesizer.write() keeps emitting audio buffers that get scheduled into the 
next utterance, so a fragment of the old speech tails into the new one.

The fix (which I've been running daily for weeks with no more out-of-order 
speech):

- On stop, do a selective flush of the pending queue — drop the speech/audio 
items but keep the queued settings commands (voice, rate, punctuation). A 
full flush breaks things: you lose the punctuation command and start hearing 
"dot," "comma," etc.
- Add a generation counter: every stop bumps it, and buffers from an older 
generation are dropped in the write callback. That kills the in-flight leak.

(The same patched build also splits chunks on sentence/clause boundaries 
instead of a raw ~15-word count, which fixed a separate mid-sentence-prosody 
problem — but the two fixes above are the ones that address your out-of-order 
issue specifically.)

My friendly AI assistant has suggested that I get these upstream to 
robertmeta, but I've never opened a pull request before, so I'm not sure of 
the right etiquette here, or how I would even do it, though it looks easy 
enough to figure out. So I'll put it to you: would a PR be welcome, or is it 
better to just email the patch/diff directly? Happy to do either. And Parham, 
I'm glad to send you the diff off-list in the meantime if you want to try it 
or look at the approach. 

Best,
Andrew