Fix #2839: initialize punc_res before conditional to prevent UnboundLocalError

[Lidang-Jiang]

Summary

Fix UnboundLocalError when punc_model is None or not provided in AutoModel.generate() with VAD enabled.

Root cause: In inference_with_vad(), punc_res is only assigned inside the if self.punc_model is not None: block, but downstream code paths (punc_segment speaker diarization and sentence_timestamp) access it unconditionally.

Fix:

• Initialize punc_res = None before the conditional block
• Add punc_res is None guards with appropriate log messages in the two affected code paths

Before (UnboundLocalError on unpatched code)

$ python -m pytest tests/test_punc_model_none.py -v

tests/test_punc_model_none.py::TestPuncModelNone::test_punc_model_none_basic PASSED [ 33%]
tests/test_punc_model_none.py::TestPuncModelNone::test_punc_model_with_value_still_works PASSED [ 66%]
tests/test_punc_model_none.py::TestPuncModelNone::test_sentence_timestamp_with_punc_model_none FAILED [100%]

________ TestPuncModelNone.test_sentence_timestamp_with_punc_model_none ________

E                       UnboundLocalError: cannot access local variable 'punc_res' where it is not associated with a value

funasr/auto/auto_model.py:677: UnboundLocalError

=========================== short test summary info ============================
FAILED tests/test_punc_model_none.py::TestPuncModelNone::test_sentence_timestamp_with_punc_model_none
======================== 1 failed, 2 passed in 4.10s ==========================

After (all tests pass)

$ python -m pytest tests/test_punc_model_none.py -v

============================= test session starts ==============================
platform linux -- Python 3.12.12, pytest-9.0.2, pluggy-1.6.0
rootdir: /home/devuser/workspace/FunASR-2839
plugins: hydra-core-1.3.2
collecting ... collected 3 items

tests/test_punc_model_none.py::TestPuncModelNone::test_punc_model_none_basic PASSED [ 33%]
tests/test_punc_model_none.py::TestPuncModelNone::test_punc_model_with_value_still_works PASSED [ 66%]
tests/test_punc_model_none.py::TestPuncModelNone::test_sentence_timestamp_with_punc_model_none PASSED [100%]

============================== 3 passed in 3.04s ===============================

Test plan

• test_punc_model_none_basic — basic ASR with punc_model=None returns correct text
• test_sentence_timestamp_with_punc_model_none — sentence_timestamp=True with punc_model=None returns empty sentence_info instead of crashing
• test_punc_model_with_value_still_works — normal flow with punc_model provided still produces punctuated text

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request fixes an UnboundLocalError in the inference_with_vad method by ensuring punc_res is properly initialized and checked before use when punc_model is None. The changes also include code formatting improvements and a new test suite to verify the fix. Feedback suggests refining the docstring in the new test file to accurately reflect that only the punc_model=None case is tested, as empty strings are not covered and may lead to different initialization issues.

[gemini-code-assist]

The docstring states that these tests cover punc_model=None or an empty string. However, the tests only cover the None case. An empty string for punc_model would likely cause a different failure mode during model initialization, not the UnboundLocalError this PR aims to fix. To avoid confusion, please update the docstring to only mention the punc_model=None case.

Suggested change

	"""Tests for issue #2839: punc_model=None or empty string should not cause UnboundLocalError."""
	"""Tests for issue #2839: punc_model=None should not cause UnboundLocalError."""