fix(runtime): cut long hotwords to 10 characters and skip empty ones

runtime/python/libtorch/funasr_torch/paraformer_bin.py

@@ -365,7 +365,10 @@ def __call__(
         return asr_res
 
     def proc_hotword(self, hotwords):
-        hotwords = hotwords.split(" ")
+        # Like the C++ runtime, skip empty hotwords and keep the first max_hotword_len
+        # characters of longer ones: the exported embedding model takes that many per hotword.
+        max_hotword_len = 10
+        hotwords = [hotword[:max_hotword_len] for hotword in hotwords.split()]
         hotwords_length = [len(i) - 1 for i in hotwords]
         hotwords_length.append(0)
         hotwords_length = np.array(hotwords_length)
@@ -385,7 +388,7 @@ def word_map(word):
 
         hotword_int = [word_map(i) for i in hotwords]
         hotword_int.append(np.array([1]))
-        hotwords = pad_list(hotword_int, pad_value=0, max_len=10)
+        hotwords = pad_list(hotword_int, pad_value=0, max_len=max_hotword_len)
         return torch.tensor(hotwords), hotwords_length
 
     def bb_infer(

runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py

@@ -407,7 +407,10 @@ def __call__(
         return asr_res
 
     def proc_hotword(self, hotwords):
-        hotwords = hotwords.split(" ")
+        # Like the C++ runtime, skip empty hotwords and keep the first max_hotword_len
+        # characters of longer ones: the exported embedding model takes that many per hotword.
+        max_hotword_len = 10
+        hotwords = [hotword[:max_hotword_len] for hotword in hotwords.split()]
         hotwords_length = [len(i) - 1 for i in hotwords]
         hotwords_length.append(0)
         hotwords_length = np.array(hotwords_length)
@@ -428,7 +431,7 @@ def word_map(word):
         hotword_int = [word_map(i) for i in hotwords]
 
         hotword_int.append(np.array([1]))
-        hotwords = pad_list(hotword_int, pad_value=0, max_len=10)
+        hotwords = pad_list(hotword_int, pad_value=0, max_len=max_hotword_len)
 
         return hotwords, hotwords_length
 

tests/test_runtime_eval_native.py

@@ -310,3 +310,32 @@ def test_onnx_wrappers_read_non_wav_paths(native_fixture, tmp_path, monkeypatch,
     model = funasr_onnx.Paraformer(str(directory), device_id=DEVICE, intra_op_num_threads=1)
     assert model(str(audio)) == model(expected)
     assert {path.name: path.read_bytes() for path in tmp_path.iterdir()} == files
+
+
+@pytest.mark.parametrize("backend", ["torch", "onnx"])
+@pytest.mark.parametrize("wrapper", ["ContextualParaformer", "SeacoParaformer"])
+def test_hotword_wrappers_skip_empty_and_cut_long_hotwords(
+    hotword_fixture, monkeypatch, backend, wrapper
+):
+    directory, paths = hotword_fixture
+    location = "libtorch" if backend == "torch" else "onnxruntime"
+    monkeypatch.syspath_prepend(str(ROOT / "runtime/python" / location))
+    module = importlib.import_module("funasr_" + backend + ".paraformer_bin")
+    model = getattr(module, wrapper)(str(directory), intra_op_num_threads=1)
+    # Like the C++ runtime: skip empty hotwords, and keep the first 10 characters of a longer
+    # one (the exported model_eb takes 10 per hotword). Ids: <s> 1, 你 3, 好 4, 魔 5, 搭 6.
+    long_hotwords = " 魔搭  你好魔搭你好魔搭你好魔搭 "
+    matrix, positions = model.proc_hotword(long_hotwords)
+    assert matrix.tolist() == [
+        [5, 6, 0, 0, 0, 0, 0, 0, 0, 0],
+        [3, 4, 5, 6, 3, 4, 5, 6, 3, 4],
+        [1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+    ]
+    assert positions.tolist() == [1, 9, 0]
+    matrix, positions = model.proc_hotword("")
+    assert matrix.tolist() == [[1, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
+    assert positions.tolist() == [0]
+    # Both decode every file; the 12-character hotword used to stop the call with a ValueError.
+    for hotwords in (long_hotwords, ""):
+        result = model(paths, hotwords)
+        assert [row["preds"][0] for row in result] == ["你" * 4, "你" * 11, "你" * 17]