fix: guard neighbour lookups in replace_blank (IndexError on trailing space)

[koriyoshi2041]

replace_blank (src/voxcpm/utils/text_normalize.py) crashes on text ending in a space, and mishandles text starting with a space.

for i, c in enumerate(text):
    if c == " ":
        if (text[i + 1].isascii() and text[i + 1] != " ") and (text[i - 1].isascii() and text[i - 1] != " "):
            out_str.append(c)
    else:
        out_str.append(c)

The intent is to keep a space only when it sits between two ASCII word characters, but the neighbour lookups aren't bounds-checked:

• A trailing space makes text[i + 1] raise IndexError: string index out of range. replace_blank("hello ") crashes. This is reachable from TextNormalizer.normalize() for any zh text that ends in a space.
• A leading space makes text[i - 1] read text[-1] (the last character), so the decision wraps around and a leading space can be wrongly kept.

Fix: only keep the space at interior positions (0 < i < len(text) - 1), so edge spaces are dropped (they aren't between two words) and no out-of-range access happens. Interior spacing (e.g. "a b" kept, "中 文" → "中文") is unchanged.

Added tests/test_text_normalize.py covering trailing/leading/interior spaces and a CJK-adjacent space.

Note: I verified the function's behaviour directly (it's pure string logic), but couldn't run the full test locally because importing the package pulls in torch/torchaudio and a native torchaudio lib fails to load on my machine — CI should exercise the added test normally.

[a710128]

Thanks for the clean fix! 🙏

I pulled the branch down and reproduced both bugs against main:

• replace_blank("hello ") → IndexError: string index out of range (trailing space, text[i + 1])
• replace_blank(" hello") → ' hello' — the leading space is wrongly kept because text[i - 1] wraps around to the last character

The 0 < i < len(text) - 1 guard is the right minimal fix: edge spaces aren't between two words, so dropping them is correct, and interior spacing is unchanged.

I also ran your test file locally (had to pip install -e . first, which pulls torch — that's likely why it didn't run on your machine), and all 3 tests pass. I spot-checked a few extra cases too, all behaving as expected:

• "a b" → "ab", "中 a" → "中a", "a 中" → "a中", "" / " " / " " all safe

LGTM, approving. 🚀

Optional (non-blocking): a couple more assertions like "a b" → "ab" and CJK/ASCII mixed cases would lock down the semantics nicely, but the current coverage is already solid.

[a710128]

Thank you for catching this and for the clear write-up — your root-cause analysis (trailing-space IndexError + leading-space negative-index wrap-around) is spot on, and the fix is correct.

Just a heads up: this overlaps with #320, which targets the same bug. Both PRs even add a tests/test_text_normalize.py, so they'd conflict on merge. We're planning to go with #320 since its tests stub the heavy dependencies and run in isolation (no torch/torchaudio import chain — which is also why you couldn't run them locally here), with a bit broader coverage.

So we'll close this one as a duplicate, but your report and fix were genuinely valuable — thanks for taking the time to dig in and explain it so clearly. 🙏

[koriyoshi2041]

Thanks for checking this and for the detailed note. #320 covers the same boundary-space bug with broader isolated tests, so closing this one to keep the queue clean.