Injecting or disrupting cadence in tts output

[chris-english]

Single spaces in text to be transformed to speech can have two impacts that may result from this. Under voice clone, pitch might increase due to the 'time of a space' aspect when translating to speech.
(This can be adjusted with ffmpeg if built with rubberband, though this query hopes to avoid this additional post-processing step.)
The uniformity of single spacing between words also seems unnatural to the ear. Several experiments suggest that one can only increase the number of spaces between words from 1 -> 2 on the text to be processed side. At 3 and above, the output becomes chaotic (single letters pronounced interspersed between whole word produced). The following work on the text to be sent through tts process, limited to changing from 1->2 spaces in some instances:

cadence_for2 <- (sent) { # sentence(s), something happens on 'some' of the spaces in, a paragraph
space_where <- unlist(gregexpr(' ', sent) )
sent_split <- unlist(strsplit(sent, split = ''))
space_mod3 <- space_where[which(space_where %% 3 == 0)]
for (k in 1:length(space_mod3) ) {
sent_split[space_mod3[k]] <- sub(' ', ' ', sent_split[space_mod3[k]]) # swap to 2 spaces
}
sent_for2 <- paste0(sent_split, collapse = '')
return(sent_for2)
}
then use returned text object in tts_chunked

cadence2 <- (sent) { # sentence(s), a replacement is possible upon every space encountered
spaces <- vector(mode = 'character', length = 2)
spaces <- c(' ',' ') # 1 or 2
space_where <- unlist(gregexpr(' ', sent) )
num_space <- length(unlist(gregexpr(' ', sent) ) )
set.seed(42)
inter <- suppressWarnings(rbind(unlist(strsplit(sent, split = ' ') ), sample(spaces, num_space, replace = TRUE) ) )
attributes(inter) <- NULL # interleave, pads ending, useful tts_chunked for between sentence space
sent_cad2 <- paste0(inter, collapse = '')
return(sent_cad2)
}
then use returned text object in tts_chunked

Are there model side variants of the special character [SPACE] that could be invoked to introduce more variability in interword spacing?

[TroyHernandez]

Thank you Chris. I appreciate your patience.

All of your testing and 3 submitted issues prompted me to do a full top-to-bottom review against the Python source (0.1.4). The LLM agents and I found a bunch of things that needed fixing before CRAN submission. The situation is greatly improved because of your work.

On #5 specifically, the 1→2-spaces-works, 3+-breaks observation was
specific enough to chase down properly, and it ended up exposing a real
divergence from upstream. I appreciate how much careful experimentation you
packed into this report.

What's actually happening:

• [SPACE] is a single token (id 2) in the vocab, and no BPE merges involve
  it. So n spaces become n [SPACE] tokens; there's no double-space token
  and no "variant" space tokens to invoke.
• Upstream Python collapses ALL whitespace runs to single spaces in
  punc_norm() before tokenizing. Your double-space trick never reaches the
  model there. It only did something in this package because we were
  skipping punc_norm on the standard path (a port bug, fixed in Python fidelity: full parity review fixes (32 findings) #6).
• That also explains the 3+ spaces chaos: repeated [SPACE] tokens are
  out-of-distribution, and isolated letters between spaces read like
  initialisms to the model, hence the letter-by-letter pronunciation.

So with #6 merged, space injection is normalized away (matching upstream),
and the supported cadence levers are:

1. Punctuation. Commas insert pauses; upstream deliberately rewrites
   ..., ;, : into commas for exactly this reason. Sentence-final
   ./?/! shape phrasing.
2. Paralinguistic tokens — this is the direct answer to your closing
   question. The vocab has a bank of special tokens the model was trained
   on: [UH], [UM], [giggle], [laughter], [inhale], [exhale],
   [sigh], [gasp], [groan], [whisper], [mumble], [sniff],
   [cough], [clear_throat], [shhh], [singing], [humming], and
   more. Before Python fidelity: full parity review fixes (32 findings) #6 the R tokenizer spelled these out letter by letter;
   they now tokenize correctly, so [inhale] or [UH] mid-sentence
   inserts a natural pause/disfluency. [sigh] before a clause slows the
   onset nicely.
3. exaggeration changes pacing and energy, and lower cfg_weight
   tends to slow delivery.

Would love to hear how the paralinguistic tokens work for your use case.