[MohammadHijjawi97]

Summary

python -m funasr.metrics.wer crashes when any utterance in the reference file has an empty transcript (for example a silent clip written as just utt2). print_cer_detail divides the per-utterance correct/error counts by the number of reference words, which is 0 for that line. The ZeroDivisionError aborts the run before the %WER / %SER summary is written, so a single empty reference loses the whole report.

This reports that utterance's rates against one word instead. The counts in the detail line, and the overall WER/SER, are unchanged.

Reproduction:

# ref.txt                 # hyp.txt
utt1 the cat sat          utt1 the cat sat
utt2                      utt2 uh
utt3 good morning         utt3 good morning

python -m funasr.metrics.wer ++ref_file=ref.txt ++hyp_file=hyp.txt ++cer_file=out.wer

Before:

  File "funasr/metrics/wer.py", line 222, in print_cer_detail
    + "{:.2%}".format(rst["cor"] / rst["nwords"])
ZeroDivisionError: division by zero

out.wer stops after utt1, with no summary.

After (out.wer):

utt2(nwords=0,cor=0,ins=1,del=0,sub=0) corr:0.00%,cer:100.00%
...
%WER 20.0 [ 1 / 5, 1 ins, 0 del, 0 sub ]
%SER 33.33 [ 1 / 3 ]
Scored 3 sentences, 0 not present in hyp.

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

• Runtime/deployment command tested

• pytest tests/test_metrics_wer.py: 2 passed; both fail with ZeroDivisionError without the fix. No model or audio is needed.

• pytest tests/test_metrics_common_rapidfuzz.py: passes.

• Ran python -m funasr.metrics.wer on the files above before and after the change (Windows 11 / Python 3.13).

• black --line-length=100 (24.4.0, as pinned in pre-commit) on the changed lines and the new test.

User impact

Anyone scoring a model with funasr.metrics.wer (as in examples/industrial_data_pretraining/paraformer/lora_cer.sh) on a test set containing a silent or untranscribed utterance gets a complete CER/WER report instead of a traceback.

Notes for reviewers

Only the per-utterance percentages in print_cer_detail change, and only when the reference is empty: the cer shown is then the number of inserted words x 100%. The aggregate %WER already handled this case (it divides by the total word count), so it is unchanged. The recipe copies under examples/aishell/*/utils/compute_wer.py have the same division and are not touched here.

[LauraGPT]

Validated exact head 9ff47d29e1007c0cb6e6677641bb6bdf46cb178b and synthetic merge 33c746408ea2482bd49b9087b41aa60e3a24d18d over protected main e5be9c9827ee139a158c9e87f893275f34ca5e9d.

• Base plus the submitted regression tests: 2 failed with the reproduced ZeroDivisionError.
• Exact head: 2/2 submitted WER regressions and 2/2 adjacent RapidFuzz tests passed.
• Synthetic merge: the same 4/4 tests passed.
• Changed-file compileall and recursive tree reconciliation passed; exactly funasr/metrics/wer.py and tests/test_metrics_wer.py differ from main.

This is a CPU-only metrics change; no model loading, ASR inference, or GPU runtime behavior is involved. The validation is targeted, not a claim that the full repository suite ran.

The two search matches, closed PRs #3105 and #3114, changed funasr/metrics/common.py aggregate metrics; this PR fixes the separate per-utterance formatter in funasr/metrics/wer.py.