[MohammadHijjawi97]

Summary

With merge_vad=True, AutoModel.inference_with_vad read merge_length_s from the construction kwargs before the per-call options were merged in. A value passed to generate() was therefore ignored, and the construction value or the 15 s default was used instead. The SenseVoice README and examples/industrial_data_pretraining/sense_voice/demo.py both pass merge_length_s to generate(), so the documented usage had no effect for any value other than 15.

The merge step now uses the per-call value first, then the construction value, then the 15 s default. The merge_length_s row in docs/python_api.md / docs/python_api_zh.md no longer tells users to set it at construction.

Reproduction (VAD segments [[0, 4000], [5000, 9000], [10000, 14000]], generate(..., merge_vad=True, merge_length_s=10)):

• Before: segments sent to ASR [[0, 14000]] (15 s default used)
• After: [[0, 9000], [9000, 14000]]

Type of change

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

Validation

• python -m compileall funasr examples tests
• Docs or links checked (tests/test_python_api_docs_contract.py)
• Runtime/deployment command tested

New tests/test_merge_vad_length.py (CPU only, mocked models, no downloads): the two per-call cases fail on main and pass with this change; default and construction-time behavior is unchanged. tests/test_punc_model_none.py and tests/test_python_api_docs_contract.py pass.

User impact

Users following the SenseVoice examples with merge_vad=True can now control the merged segment length per call, as the README describes.

Notes for reviewers

Construction-time merge_length_s keeps working; a per-call value now takes precedence, consistent with how other generate() options are applied.

[LauraGPT]

Reviewed all four changed files at 099c2f4404be97c28ce20aa980afd8295d8dab37. Reading the per-call option before VAD merging fixes the ignored override while preserving the construction/default fallback and disabled-merge behavior. Both language references match the implementation.

Validation on the current-main candidate tree 175297234949cfbbba44e5b377f06b7d80c3b5c3:

• Baseline reproduced the two submitted override failures; all 36 selected CPU regression/documentation tests passed with the change.
• 32 real L20Z FP32 integration cases through AutoModel.generate, configuration reset, audio slicing and batched inference: baseline failed 8, candidate passed 32/32. Construction defaults/10s, repeated per-call 10s/default/15s/disabled, 1/2 inputs and 1s/300s batch budgets were covered. Each phase recorded 1,072 CUDA events and per-case raw results.
• The GPU backends are explicit analytic torch fixtures with a zero scalar parameter, not pretrained ASR/VAD weights. This establishes wrapper/CUDA integration behavior, not ASR quality, Windows/NPU support or a full-suite result. No models were downloaded.
• Exact-head hosted CI logs and NumPy 1.26.4/2.4.0 artifacts were inspected. Those existing workflows do not run the new four-case regression; its passing result above is local, not attributed to hosted CI.

The earlier stale GitHub potential merge was not used as the tested current-main tree. The candidate retains the intervening standalone runtime release changes. Source fix only: existing published package bytes are unchanged. Normal merge and current protection checks apply.