fix(frontend): stop cmvn readers from crashing on blank lines or returning empty statistics

[Lesereingrape]

Summary

Both cmvn readers in the installed package walk the file line by line and index the result without checking it: line_item[0] on a blank line raises IndexError: list index out of range, and the unbounded one-line lookahead lines[i + 1] raises the same error when a <AddShift> / <Rescale> tag is the last line. Because the parse is a pure line-matching loop, a file whose tags do not start a line on its own (which Kaldi does emit — the whole <Nnet> on one line) matched nothing and load_cmvn returned empty statistics with no error at all; the failure then only surfaced on the first inference as RuntimeError: The size of tensor a (N) must match the size of tensor b (0).

Fixes #3759

Three-line change per reader (funasr/frontends/wav_frontend.py:15, funasr/frontends/default.py:390):

lines = [line for line in f.readlines() if line.split()]          # skip blank/whitespace-only lines
line_item = lines[i + 1].split() if i + 1 < len(lines) else []     # bounded lookahead
if line_item and line_item[0] == "<LearnRateCoef>":
...
if not means_list or not vars_list:
    raise ValueError(f"No <AddShift>/<Rescale> statistics found in cmvn file: {cmvn_file}")

Skipping blank lines first (rather than guarding inside the loop) also makes <AddShift> + blank line + <LearnRateCoef> parse, which a per-line guard would silently drop. The ValueError follows the convention the sibling runtime loader already uses (runtime/python/onnxruntime/funasr_onnx/utils/frontend.py:148 names the offending cmvn file), and blank-line skipping matches how this repository reads other plain-text input (funasr/bin/realtime_ws.py:1363, funasr/utils/compute_det_ctc.py:59).

Type of change

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

Validation

Measured on main @ 66d7a4c2, Python 3.13.7 / numpy 2.5.3 / torch 2.14.1+cpu, Windows, with PYTHONPATH pointed at the worktree (verified by printing funasr.__file__):

• RED with only the two source files restored from origin/main and the new tests present: python -m pytest -q tests/test_numpy_compatibility.py -k cmvn → 2 failed, 3 passed.

• GREEN with the fix: same command → 5 passed; whole file python -m pytest -q tests/test_numpy_compatibility.py → 9 passed, including the two pre-existing cmvn tests and test_frontend_applies_cmvn_and_preserves_padding.

• Regression guard as a test of its own: test_cmvn_reader_loads_the_am_mvn_shipped_by_the_runtime asserts the am.mvn this repository ships under runtime/triton_gpu/model_repo_sense_voice_small/feature_extractor/ still loads as (2, 560) finite values — it passes on main too, so a file that parses correctly today provably does not change.

• python -m compileall -q funasr/frontends tests/test_numpy_compatibility.py → exit 0.

• Formatting: the repository is not black-clean at main (e.g. funasr/frontends/default.py docstring block at 392-395 is already reformatted by black --line-length=100), so black --check fails on both the pristine and the patched file. Checked instead that black --line-length=100 leaves every added line untouched — the only hunks it proposes in these files are pre-existing ones outside my diff.

• python -m compileall funasr examples tests

• Docs or links checked

• Runtime/deployment command tested

User impact

Paraformer / SenseVoice users who point cmvn_file at a hand-edited or concatenated .mvn file — or at a Kaldi artifact written on a single line — currently get either a bare IndexError during frontend construction, or a frontend that builds fine and fails at the first model.generate() call. Both cases now either parse or raise an error that names the file.

Notes for reviewers

• The same loop is duplicated in five more places under runtime/ (runtime/python/libtorch/, runtime/python/onnxruntime/, runtime/triton_gpu/, export_lfr_cmvn_pe_onnx.py). Those are standalone export/deployment helpers rather than the installed package, so this PR deliberately leaves them alone; the issue lists them.
• Two of the three possible failure modes were silent or crash-only, so the ValueError is a behaviour change in one narrow case: a cmvn file that yielded empty statistics used to return them. Such a frontend could never produce features (it hit the RuntimeError above on the first call), so this only converts a downstream failure into an actionable one.
• The practical limitation to note: a fork PR in this repository does not trigger workflow runs, so all evidence above is local, and the reproduction needs no model weights or downloads.
• Prepared with an AI coding agent; a human maintainer at the PR author's side has not re-reviewed the diff beyond the checks listed above.

[LauraGPT]

Reviewed exact head d0b9cc9 against current-main integration. New CMVN cases first fail on old runtime (2 failures / 3 passes), then all 5 selected cases pass. Independent checks cover 30 parser boundaries including empty, partial, truncated, blank/CRLF input and preserved values/dtypes. Replaying the shipped am.mvn with the reporter's blank-line edits and actual WavFrontend construction gives 6 failures / 3 passes before the fix, then 9 passes with exactly unchanged statistics. The current NumPy matrix runs pass 181 tests + 10 subtests in each lane; all 433 exported package Python files match the tested Git tree after restoring omitted setup.py in the test snapshot. Both approved head workflows now succeed; both NumPy artifacts show 123 entries, zero failures/errors/skips, exact head source and all three added regressions. The only main advance since runtime validation is #3764's independently validated SDK image links. No blocking findings. Compact single-line CMVN is rejected early, not newly supported; five standalone runtime/export helper copies remain unchanged. Evidence is Linux/Python3.12/CPU, not Windows hardware or checkpoint inference.