Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fix(runtime): handle blank and incomplete CMVN files

BillVan1945b04804c96dcf
diff --git a/runtime/python/libtorch/funasr_torch/utils/frontend.py b/runtime/python/libtorch/funasr_torch/utils/frontend.py
index 743ee5e5f..2772ad95d 100644
--- a/runtime/python/libtorch/funasr_torch/utils/frontend.py
+++ b/runtime/python/libtorch/funasr_torch/utils/frontend.py
@@ -125,25 +125,29 @@ class WavFrontend:
self,
) -> np.ndarray:
with open(self.cmvn_file, "r", encoding="utf-8") as f:
- lines = f.readlines()
+ lines = [line for line in f.readlines() if line.split()]
means_list = []
vars_list = []
for i in range(len(lines)):
line_item = lines[i].split()
if line_item[0] == "<AddShift>":
- line_item = lines[i + 1].split()
- if line_item[0] == "<LearnRateCoef>":
+ line_item = lines[i + 1].split() if i + 1 < len(lines) else []
+ if line_item and line_item[0] == "<LearnRateCoef>":
add_shift_line = line_item[3 : (len(line_item) - 1)]
means_list = list(add_shift_line)
continue
elif line_item[0] == "<Rescale>":
- line_item = lines[i + 1].split()
- if line_item[0] == "<LearnRateCoef>":
+ line_item = lines[i + 1].split() if i + 1 < len(lines) else []
+ if line_item and line_item[0] == "<LearnRateCoef>":
rescale_line = line_item[3 : (len(line_item) - 1)]
vars_list = list(rescale_line)
continue
+ if not means_list or not vars_list:
+ raise ValueError(
+ f"No <AddShift>/<Rescale> statistics found in cmvn file: {self.cmvn_file}"
+ )
means = np.array(means_list).astype(np.float64)
vars = np.array(vars_list).astype(np.float64)
cmvn = np.array([means, vars])
diff --git a/runtime/python/onnxruntime/funasr_onnx/utils/frontend.py b/runtime/python/onnxruntime/funasr_onnx/utils/frontend.py
index f150ccbf8..4218718b6 100644
--- a/runtime/python/onnxruntime/funasr_onnx/utils/frontend.py
+++ b/runtime/python/onnxruntime/funasr_onnx/utils/frontend.py
@@ -142,24 +142,26 @@ def load_cmvn(cmvn_file: Union[str, Path]) -> np.ndarray:
raise FileNotFoundError("cmvn file not exits")
with open(cmvn_file, "r", encoding="utf-8") as f:
- lines = f.readlines()
+ lines = [line for line in f.readlines() if line.split()]
means_list = []
vars_list = []
for i in range(len(lines)):
line_item = lines[i].split()
if line_item[0] == "<AddShift>":
- line_item = lines[i + 1].split()
- if line_item[0] == "<LearnRateCoef>":
+ line_item = lines[i + 1].split() if i + 1 < len(lines) else []
+ if line_item and line_item[0] == "<LearnRateCoef>":
add_shift_line = line_item[3 : (len(line_item) - 1)]
means_list = list(add_shift_line)
continue
elif line_item[0] == "<Rescale>":
- line_item = lines[i + 1].split()
- if line_item[0] == "<LearnRateCoef>":
+ line_item = lines[i + 1].split() if i + 1 < len(lines) else []
+ if line_item and line_item[0] == "<LearnRateCoef>":
rescale_line = line_item[3 : (len(line_item) - 1)]
vars_list = list(rescale_line)
continue
+ if not means_list or not vars_list:
+ raise ValueError(f"No <AddShift>/<Rescale> statistics found in cmvn file: {cmvn_file}")
means = np.array(means_list).astype(np.float64)
vars = np.array(vars_list).astype(np.float64)
cmvn = np.array([means, vars])
diff --git a/tests/test_runtime_cmvn.py b/tests/test_runtime_cmvn.py
new file mode 100644
index 000000000..97889e2fc
--- /dev/null
+++ b/tests/test_runtime_cmvn.py
@@ -0,0 +1,105 @@
+"""Exercise the standalone runtime frontends without model weights or engines."""
+
+import importlib.util
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+
+ROOT = Path(__file__).resolve().parents[1]
+CMVN = "<AddShift> 2 2\n<LearnRateCoef> 1 [ 1 -2 ]\n<Rescale> 2 2\n<LearnRateCoef> 1 [ 0.5 2 ]\n"
+
+
[email protected](params=["onnxruntime/funasr_onnx", "libtorch/funasr_torch"])
+def frontend_class(request):
+ # Loading the frontend directly keeps these tests independent of the
+ # optional inference engines and exercises each package's real constructor.
+ path = ROOT / "runtime/python" / request.param / "utils/frontend.py"
+ spec = importlib.util.spec_from_file_location("runtime_cmvn_frontend", path)
+ module = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(module)
+ return module.WavFrontend
+
+
+def test_valid_cmvn_preserves_statistics_and_normalized_features(frontend_class, tmp_path):
+ path = tmp_path / "valid.mvn"
+ path.write_text(CMVN, encoding="utf-8")
+ frontend = frontend_class(cmvn_file=str(path))
+
+ assert frontend.cmvn.dtype == np.float64
+ np.testing.assert_array_equal(frontend.cmvn, [[1, -2], [0.5, 2]])
+ features = np.array([[0, 2], [-2, 4]], dtype=np.float32)
+ np.testing.assert_array_equal(frontend.apply_cmvn(features), [[0.5, 0], [-0.5, 4]])
+
+
[email protected]("placement", ["leading", "trailing", "between_sections", "after_tags"])
+def test_cmvn_accepts_blank_lines(frontend_class, tmp_path, placement):
+ blank = " \t\n\n"
+ if placement == "leading":
+ content = blank + CMVN
+ elif placement == "trailing":
+ content = CMVN + blank
+ elif placement == "between_sections":
+ content = CMVN.replace("<Rescale>", blank + "<Rescale>")
+ else:
+ content = CMVN.replace("<LearnRateCoef>", blank + "<LearnRateCoef>")
+ path = tmp_path / "blank lines.mvn"
+ path.write_text(content, encoding="utf-8")
+
+ frontend = frontend_class(cmvn_file=str(path))
+
+ np.testing.assert_array_equal(frontend.cmvn, [[1, -2], [0.5, 2]])
+
+
+ "content",
+ [
+ "",
+ " \t\n",
+ "<Nnet>\n</Nnet>\n",
+ "<AddShift> 2 2\n",
+ "<Rescale> 2 2\n",
+ "<AddShift> 2 2\n<LearnRateCoef> 1 [ 1 -2 ]\n",
+ "<Rescale> 2 2\n<LearnRateCoef> 1 [ 0.5 2 ]\n",
+ CMVN.replace("\n", " "),
+ ],
+ ids=[
+ "empty",
+ "blank",
+ "no_stats",
+ "truncated_shift",
+ "truncated_scale",
+ "shift_only",
+ "scale_only",
+ "compact",
+ ],
+)
+def test_incomplete_cmvn_fails_during_construction(frontend_class, tmp_path, content):
+ path = tmp_path / "incomplete.mvn"
+ path.write_text(content, encoding="utf-8")
+
+ with pytest.raises(ValueError) as error:
+ frontend_class(cmvn_file=str(path))
+
+ assert str(path) in str(error.value)
+
+
+def test_shipped_cmvn_is_unchanged_by_blank_lines(frontend_class, tmp_path):
+ shipped = ROOT / "runtime/triton_gpu/model_repo_sense_voice_small/feature_extractor/am.mvn"
+ original = frontend_class(cmvn_file=str(shipped), lfr_m=7, lfr_n=6, dither=0.0)
+ path = tmp_path / "shipped-with-blanks.mvn"
+ path.write_text("\n" + "\n \t\n".join(shipped.read_text().splitlines()) + "\n\n")
+ edited = frontend_class(cmvn_file=str(path), lfr_m=7, lfr_n=6, dither=0.0)
+
+ assert original.cmvn.shape == (2, 560)
+ assert np.isfinite(original.cmvn).all()
+ np.testing.assert_array_equal(edited.cmvn, original.cmvn)
+ # Real filterbank + LFR + CMVN on a deterministic waveform, not a model stub.
+ waveform = np.sin(2 * np.pi * 440 * np.arange(1600) / 16000).astype(np.float32) * 0.1
+ original_features, _ = original.fbank(waveform)
+ edited_features, _ = edited.fbank(waveform)
+ original_features, original_length = original.lfr_cmvn(original_features)
+ edited_features, edited_length = edited.lfr_cmvn(edited_features)
+ assert edited_length == original_length
+ np.testing.assert_array_equal(edited_features, original_features)