[LauraGPT]

Hotword Scope Correction

PR #3798 fixes cross-script fuzzy hotword misreplacement, not local hotword-file loading. In particular, Chinese-only text such as 这个端口 must not acquire the C prefix from the mixed-script hotword C端; complete mixed-script homophones remain supported. The earlier description was incorrect.

Nine additional checks of the final installed 1.4.18 wheel passed with real pypinyin/rapidfuzz dependencies: the reporter's AI/digit example, Chinese-only prefix rejection, unchanged valid text, and complete mixed-script homophone controls. These are post-publication CPU text checks, not file-parser coverage, new GPU/model inference or a full suite.

The immutable v1.4.18 tag, wheel/sdist and their packaged README/PyPI description retain the original erroneous summary sentence. This erratum corrects that description; no package bytes, tag or hashes have been replaced. The runtime fix itself is present and the official PyPI bytes remain equal to the accepted GitHub packages.

FunASR 1.4.18

Source: 5315edd5db957ffd5c0035ac5b6b432560e88608 / tag v1.4.18.

Fixes

• Prevent cross-script fuzzy hotword misreplacement (fix(hotwords): prevent missing Latin prefixes in mixed-script correction #3798), while retaining complete mixed-script homophones.
• Keep timestamps aligned when special tokens are filtered (fix(postprocess): keep word timestamps aligned after filtered tokens #3802/fix: align filtered token timestamps and reject incomplete input #3803). A retained token without a timestamp now raises ValueError instead of silently shifting the mapping; an untimed trailing filtered EOS remains supported.
• Preserve Chinese VAD chunk token boundaries for sentence text, timestamps and speaker assignment (fix: preserve VAD chunk token alignment in sentence timestamps #3804).

The auxiliary ONNX/libtorch source corrections are in this Git tag; this release does not publish new independent versions of those Python packages. Reporter-specific original-recording confirmation remains outstanding.

Verification Scope

• 137 release/docs/package/runtime-contract checks passed; 2 local browser checks skipped because their browser tools were absent. Exact-head hosted build, browser and all other checks passed separately, as did the new-main checks.
• 52 affected regression cases passed with a mix of installed imports and explicit source-file loaders; these are not all installed-wheel tests. An additional 15 cases explicitly verified imports from the installed candidate wheel: 3 existing postprocessing hotword-parser controls (not tests of the cross-script matcher repair), 4 Paraformer hotword-file controls and 8 filtered-token timestamp cases. On an idle NVIDIA L20Z, the installed wheel passed 6 FP32/BF16 CUDA CIF timestamp pipelines and 8 VAD/speaker/raw-text alignment cases. Inputs and model responses are synthetic fixtures, not pretrained ASR or original-recording quality evaluation.
• Newly downloaded final-tag wheel and sdist match the accepted source files; metadata, all 13 entrypoints, strict Twine, offline installation, actual package import and CLI checks passed.
• All 12 payloads passed full SHA-256/size and archive-integrity checks. The 2 Python packages were newly downloaded from this release. The 10 unchanged runtime archives reuse complete verified cache bytes, freshly rehashed against both the new release asset metadata and runtime-llamacpp-v0.2.6. They were not downloaded anew or rebuilt. Platform binaries were not executed.
• SHA256SUMS covers all 12 payloads. This is scoped release validation, not a full repository suite or the deferred Nano nine-dataset benchmark.

GitHub release assets and the manifest are accepted. PyPI is the separate last publication and official-download verification gate.

Historical Pre-Release Plan

The dated plan below is retained history, not current release status.

Summary

Release the three normally merged post-v1.4.17 fixes: cross-script fuzzy hotword misreplacement (#3798), filtered-token timestamp alignment (#3803, retaining #3802), and Chinese VAD sentence/timestamp/speaker alignment (#3804).

Only version, release contract and multilingual release/install documentation change here. A retained token missing its timestamp now raises ValueError instead of silently shifting alignment; trailing filtered EOS remains valid. Auxiliary ONNX/libtorch source fixes do not imply new independent package versions.

Validation

• Version contract RED: 3 failures and 1 passing control before version/docs changes.
• Release/docs/package/runtime-contract suite: 137 passed, 2 local browser skips. Exact-head hosted CI, including browser, must pass before normal merge.
• Candidate wheel/sdist: exact packaged source equality, 13 entrypoints, strict Twine, offline installation, installed import and CLI passed.
• 52 affected cases passed using mixed installed imports and explicit source-file loaders; these are not all installed-wheel tests. An additional 4 hotword and 8 filtered-token timestamp cases explicitly verified imports from the installed candidate package. The installed wheel also passed 6 real L20Z FP32/BF16 synthetic CIF pipelines and 8 VAD alignment cases. These are scoped regressions, not pretrained ASR/original-recording quality or a full suite.
• Existing ten-platform runtime pointer is unchanged. Final tag Python bytes and all twelve payload digests will be separately verified; PyPI is last. No protection change or administrator bypass.

No tag, formal Release or PyPI upload is claimed by this PR. Issues remain open for reporter confirmation. Nano leaderboard work stays deferred.