[MohammadHijjawi97]

Summary

KwsCtcPrefixDecoder splits keywords on commas. A trailing, leading or doubled comma (keywords="小云小云,"), or an entry made only of punctuation, produces a keyword with no token ids. An empty token sequence matches every hypothesis, so:

• silence is reported as a detection of the empty keyword with score 1.0 (detected 1.0);
• a hypothesis that contains only part of the real keyword raises IndexError: tuple index out of range in is_sublist.

This change skips keywords that map to no tokens and logs a warning. Real keywords are decoded exactly as before.

Before (token list <blank> <unk> 小 云, keywords "小云小云,"):

silence posteriors      -> (True, '', 1.0)
only "小" then blanks   -> IndexError: tuple index out of range

After:

silence posteriors      -> (False, None, None)
only "小" then blanks   -> (False, None, None)
"小云" "小云"           -> (True, '小云小云', 1.0)

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

• python -m compileall on the changed files
• Docs or links checked
• Runtime/deployment command tested

New regression test tests/test_kws_utils.py::KwsUtilsTest::test_empty_keyword_entries_do_not_produce_detections uses synthetic CTC posteriors (no model download). It fails on main (4 failed subtests: (True, '', 1.0) != (False, None, None)) and passes with this change. tests/test_kws_utils.py, tests/test_kws_optional_output.py, tests/test_kws_streaming_continuity.py and tests/test_kws_docs.py show no new failures.

User impact

Users of the FSMN/SANM KWS models who pass several keywords as a comma-separated string no longer get false detections or crashes from an extra comma.

Notes for reviewers

The fix is in the keyword parsing only; the beam search and is_sublist are unchanged. Keywords with no tokens are skipped with a warning rather than raising, so an existing configuration with a stray comma keeps working.

[LauraGPT]

Independent agent-assisted review of external head d1ef17d4fd18f02505063e71c8e57090a0798901 against current main 940b0d2463141a2e9ace8a9fd965b4aea93f2db5, tested potential-merge tree 45a85334e8896db83d3e62224bf7edb593a0fb0d.

I read both changed files, the full decoder implementation, and all current PR discussion (none). No blocking issue found: the constructor drops only entries whose token sequence is empty; valid keyword parsing and beam search are unchanged. All 4,153 source blobs match the exact merge export, with only the two declared files different from main.

Verification on Python 3.12.3 / PyTorch 2.8.0a0+5228986c39.nv25.06:

• The new regression against the unchanged decoder fails with the reported (True, '', 1.0) silence detection.
• Exact merge source: all 45 tests in test_kws_utils, test_kws_optional_output, test_kws_streaming_continuity and test_kws_docs pass, with no skips.
• Real NVIDIA L20Z CUDA 12.9 execution: 48 CPU/CUDA comparisons (96 decode calls) across FP32/BF16, valid/leading/trailing/doubled-comma/punctuation/all-empty configurations, silence/partial/full-keyword posteriors. Expected detections and scores agree in every case. Softmax runs on the selected device; the production decoder then transfers its posterior tensor to CPU as designed.
• Exact-head hosted kws-output check is successful.

These are focused KWS tests and synthetic CTC inputs, not pretrained acoustic-model inference, audio-quality evaluation or the full repository suite. No model download or reference/threshold modification. Approval is from a different account than the contributor and does not bypass branch protection.