Bug: cmvn readers raise IndexError on blank lines and return empty statistics without error

[Lesereingrape]

🐛 Bug

FunASR has two cmvn readers in the installed package, and both walk the file line by line while indexing the result without any check:

• funasr/frontends/wav_frontend.py:15 load_cmvn — used by WavFrontend / WavFrontendOnline, i.e. the frontend of the Paraformer and SenseVoice models most people load first;
• funasr/frontends/default.py:390 MultiChannelFrontend._load_cmvn — used by MultiChannelFrontend (emotion2vec / campplus).

    for i in range(len(lines)):
        line_item = lines[i].split()
        if line_item[0] == "<AddShift>":          # blank line -> IndexError
            line_item = lines[i + 1].split()      # tag on the last line -> IndexError
            if line_item[0] == "<LearnRateCoef>":

Two distinct symptoms come out of this, both measured on main @ 66d7a4c2:

cmvn_file content	wav_frontend.load_cmvn	MultiChannelFrontend._load_cmvn
runtime/triton_gpu/model_repo_sense_voice_small/feature_extractor/am.mvn (a real artifact this repository ships)	OK, shape (2, 560)	OK, shape (560,)
the same file plus one trailing blank line	IndexError: list index out of range	IndexError: list index out of range
a blank (or whitespace-only) line between <AddShift> and its <LearnRateCoef> line	IndexError: list index out of range	IndexError: list index out of range
<AddShift> 2 2 as the last line (truncated file)	IndexError: list index out of range	IndexError: list index out of range
statistics present but the tags do not start a line on their own, e.g. the whole <Nnet> on one line	returns a tensor of shape (2, 0) — no error	returns two empty arrays — no error

The first four crash while the frontend is being constructed, so AutoModel(...) or a training run dies with a bare IndexError that never mentions the cmvn file. The fifth is worse because it is silent: load_cmvn returns empty statistics, WavFrontend builds without complaint, and the failure only appears at the first inference.

To Reproduce

1. Install with: pip install -e . (source checkout of main @ 66d7a4c2)
2. Copy the cmvn artifact the repository already ships and append a blank line:
   cp runtime/triton_gpu/model_repo_sense_voice_small/feature_extractor/am.mvn /tmp/am.mvn && printf '\n' >> /tmp/am.mvn
3. Run the code sample below.
4. See error: IndexError: list index out of range

Code sample

from funasr.frontends.wav_frontend import WavFrontend

WavFrontend(cmvn_file="am.mvn", fs=16000, win_length=400, hop_length=160, n_mels=80)
# IndexError: list index out of range   -- only when the file has an extra blank line;
# without the blank line the same call succeeds.

# silent variant: tags do not start a line on their own
from funasr.frontends.wav_frontend import load_cmvn
open("compact.mvn", "w").write(
    "<Nnet> <AddShift> <LearnRateCoef> 1 [ -1.0 -2.0 ] </LearnRateCoef> "
    "<Rescale> <LearnRateCoef> 1 [ 0.5 0.5 ] </LearnRateCoef> </Nnet>\n"
)
cmvn = load_cmvn("compact.mvn")   # shape (2, 0), no error

Measured with n_mels=3 on the compact file, the frontend then builds fine and the first input_feats += self.mean / apply_cmvn call fails at inference time:

RuntimeError: The size of tensor a (3) must match the size of tensor b (0) at non-singleton dimension 1

MultiChannelFrontend behaves the same way (input_feats += self.mean against an empty registered buffer, funasr/frontends/default.py:355).

Expected behavior

A cmvn file that contains valid statistics should load regardless of blank or whitespace-only separator lines, and a file that yields no statistics should raise an error naming the file instead of returning empty arrays. Files that parse correctly today must keep parsing identically — the am.mvn above is the regression case.

Error logs

  File ".../funasr/frontends/wav_frontend.py", line 27, in load_cmvn
    if line_item[0] == "<AddShift>":
IndexError: list index out of range

Environment

• OS: Windows 11 (also reproduces on Linux — the loop is platform-independent)
• Python version: 3.13.7
• FunASR version: source, main @ 66d7a4c2
• PyTorch / torchaudio version: 2.14.1+cpu / matching
• numpy: 2.5.3
• Install method: source (pip install -e .)
• Device: cpu
• No GPU, no model download needed — the reproduction uses the cmvn text file the repository already ships.

Audio details

Not applicable: no audio is involved, the defect is in the plain-text cmvn parser.

Suggested fix

For the two package readers:

1. drop blank/whitespace-only lines before pairing a section tag with the line below it — that also makes <AddShift> + blank line + <LearnRateCoef> parse, and matches how this repository already reads other plain-text inputs (funasr/bin/realtime_ws.py:1363 skips blank hotword lines, funasr/utils/compute_det_ctc.py:59 skips blank stat lines);
2. bound the one-line lookahead so a truncated file cannot raise IndexError;
3. raise a ValueError naming the file when neither <AddShift> nor <Rescale> produced statistics, instead of handing back empty arrays — the same "say which cmvn file is wrong" convention the sibling runtime loader already uses (runtime/python/onnxruntime/funasr_onnx/utils/frontend.py:148 raises FileNotFoundError("cmvn file not exits")).

Five more copies of the same loop exist under runtime/ (runtime/python/libtorch/, runtime/python/onnxruntime/, runtime/triton_gpu/, export_lfr_cmvn_pe_onnx.py); they are standalone export/deployment helpers rather than the installed package, so I would scope any PR to the two package readers.

Filed with an AI coding agent, based on the measurements above that I ran locally on this machine.