fix: prevent IndexError in Whisper word timestamp decode

[guoyangzhen]

Problem

In _split_tokens_on_unicode(), when the decoded token stream ends with a dangling Unicode replacement character (U+FFFD), the computed index can equal len(decoded_full), causing IndexError: string index out of range.

The failing line:

decoded_full[unicode_offset + decoded.index(replacement_char)] == replacement_char

When unicode_offset + decoded.index(replacement_char) >= len(decoded_full), this crashes.

Fix

Add bounds check before indexing into decoded_full:

if (
    replacement_char not in decoded
    or (unicode_offset + decoded.index(replacement_char) < len(decoded_full)
        and decoded_full[unicode_offset + decoded.index(replacement_char)] == replacement_char)
):

This ensures we only access decoded_full when the index is valid.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: whisper

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=44885&sha=a326a1

[Rocketknight1]

Your bot has created a duplicate PR at #44902 for the same issue. As we mentioned in CONTRIBUTING, we're being flooded with code agent PRs right now, so temporarily blocking you to cut down the notification spam