[BillVan1945]

Summary

Offline Fsmn_vad crashes on empty audio or recordings shorter than one analysis window. The frontend produces no feature frames, but LFR still reads the first frame, raising IndexError. In a file list, this also interrupts processing of the remaining files.

Check the actual fbank frame count before applying LFR. For zero-frame inputs, return correctly shaped empty features and skip inference. Each input keeps its empty segment list, and processing continues with the next file.

Type of change

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

Validation

Extended the existing offline VAD tests to cover empty/short inputs, analysis-window boundaries and mixed short/normal file lists. Tests verify result order, unchanged normal-audio segments and no inference calls for zero-frame inputs.

Offline/streaming VAD regressions and installed-package checks pass. Tests use the real frontend and VAD state machine with mocked neural inference.

• python -m compileall funasr examples tests
• Docs or links checked
• Runtime/deployment command tested (installed-package regression tests)

User impact

Empty or very short recordings return an empty segment list instead of aborting the call, including when mixed with normal recordings.

Notes for reviewers

The check follows the frontend configuration rather than a fixed sample count. Changes are limited to offline Fsmn_vad and its existing tests; the shared frontend and streaming VAD are unchanged.

[LauraGPT]

Reviewed both changed files at f2dfafb3848b19e13eae5e1a9f51bb682903c40e against current main 57485f744e445898cdba06d0868468bcc5086f5d. The zero-frame check uses the actual frontend length, preserves an empty independent result at each short-input position, and leaves normal LFR/scoring unchanged. The feature shape and dtype are preserved when bypassing LFR.

Maintainer validation:

• The submitted short-input regression fails on unchanged main; the failure log is retained.
• 77 selected installed-wheel CPU/package tests pass on the combined current-main tree.
• 87 newly executed L20Z/CUDA FP32 cases pass: empty/sub-frame/boundary audio, paths/arrays, mixed file lists, requested batch sizes 1/2/3 (effective batch 1), exact result ownership and inference-call counts. Actual profiles contain CUDA nodes and no CPU node fallback. This uses the existing analytic ONNX scoring/cache fixture plus real wrapper/frontend/VAD state machine, not pretrained weights or ASR-quality/full-suite validation.
• Exact-head CI logs and downloaded wheel/sdist/runtime artifacts were inspected, including actual package source bytes and source-commit markers. No CI rerun.

This is a source fix. Existing published funasr-onnx 0.4.8 is not overwritten; package delivery requires a new version. No protection bypass or approval-policy change.