[ychy7001]

Summary

Fixes #3791.

When a final call is split into two chunks and both infer() calls return [], the merge loop leaves res = {} and __call__ returns [{}]. The official runtime/python/onnxruntime/demo_paraformer_online.py checks len(rec_result) > 0 before accessing rec_result[0]["preds"][0], so this result passes the check and raises KeyError: 'preds'.

Change one runtime line to return [res] if res else []. Normal recognition and merging are unchanged, including a result from only one of the two chunks. Single-chunk results, including [{"preds": ("", [])}], are unchanged. infer() and sentence_postprocess() are untouched.

Type of change

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

Validation

• PYTHONPATH=runtime/python/onnxruntime python tests/test_funasr_onnx_paraformer_online_empty_final.py -v: 5 passed. On fbbfe01, only the both-chunks-empty test fails ([{}] != []); the other 4 pass. Tests mock feature extraction and infer() while exercising the real __call__, cache preparation and overlap splitting. No model downloads are needed.
• Official v2.0.5 FP32 ONNX export, Python 3.12.3, onnxruntime 1.20.1 CPU, numpy 1.26.4, default dither, chunk_size=[5, 10, 5]: a 9600-sample non-final silence call followed by a 16000-sample final silence call returns [{}] before the fix and triggers the demo's access error. After the fix it returns [] without an error. No overlap-slicing monkeypatch was used.
• For AISHELL-1 BAC009S0764W0129, appending 0, 0.6, 1, 1.2, 2, 3 or 5 seconds of silence and using the demo's 9600-sample step did not trigger the error. Results were identical before and after. The reproduced 16000-sample final call is larger than the demo's usual step; this does not establish that trailing silence generally crashes the demo loop.
• The new test is wired into .github/workflows/test-funasr-onnx-package.yml (path filters and a step after the wheel install).
• git diff --check passed.

Notes for reviewers

A caller could theoretically treat [{}] as a status signal, so returning [] changes that observation. The tradeoff is to represent the absence of recognition results consistently with infer() and avoid the official demo's KeyError.

Other output consistency issues, including single-chunk tuple versus split-chunk list values under preds, are outside this change. This PR does not depend on #3790. The commit contains the one-line runtime fix, a regression test file, and the workflow entry that runs the test in Validate funasr-onnx package after the wheel is installed.

[LauraGPT]

Thanks @ychy7001. Your one-line empty split-final fix and the exact five regression methods are now integrated through #3808, preserving contributor credit. After #3790 merged, only the workflow had three conflicts; the integration keeps both regressions and both event filters.

The combined tree passed 55 focused CPU tests and 20 real L20Z CUDA boundary cases. New-head and final-main Python 3.11/3.12 CI, actual wheel/sdist source bytes and both regression logs were accepted. GPU inputs/ONNX parameters are explicit synthetic fixtures, not pretrained ASR accuracy or a full suite. No new weights were downloaded.

Closing this PR as superseded, not as merged: the original fork branch/head is unchanged. The fix is on main; no new standalone ONNX PyPI release is claimed here.