From a3b21c4ef3485780c389e2eddff789df05f24b84 Mon Sep 17 00:00:00 2001 From: Tanishq Date: Mon, 9 Nov 2026 23:21:21 +0530 Subject: [PATCH] fix(onnx): load non-.wav audio paths with librosa instead of pydub load_data in the funasr_onnx Paraformer (inherited by ContextualParaformer and SeacoParaformer), streaming Paraformer, Fsmn_vad and SenseVoiceSmall sent every path that does not end in .wav through pydub.AudioSegment.from_mp3, exported the result to .wav next to the input and loaded that file. pydub is not a funasr-onnx dependency, so on a clean install every .mp3, .flac or .ogg path raised ModuleNotFoundError: No module named 'pydub', also for the en.mp3 input of demo_sencevoice_small.py. With pydub but without ffmpeg, the conversion only printed its exception and the wrapper then loaded .wav: a FileNotFoundError for a file that was never passed or, when a different .wav sat next to the input, that recording, without an error. librosa, already a dependency, reads these formats through soundfile and falls back to audioread for files libsndfile cannot decode. Drop the conversion and load every path with librosa.load, as funasr_torch and Fsmn_vad_online already do. Extend tests/test_runtime_eval_native.py: for FLAC, OGG and MP3 files next to a different .wav, the load_data of each wrapper must return librosa.load(path, sr=16000), the native-fixture Paraformer must decode the path like the loaded waveform, and no file in the directory may change. --- .../onnxruntime/funasr_onnx/paraformer_bin.py | 16 -------- .../funasr_onnx/paraformer_online_bin.py | 17 --------- .../onnxruntime/funasr_onnx/sensevoice_bin.py | 17 --------- .../python/onnxruntime/funasr_onnx/vad_bin.py | 17 --------- tests/test_runtime_eval_native.py | 37 +++++++++++++++++++ 5 files changed, 37 insertions(+), 67 deletions(-) diff --git a/runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py b/runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py index d822e73183..517252cd85 100644 --- a/runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py +++ b/runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py @@ -181,23 +181,7 @@ def plot_wave_timestamp(self, wav, text_timestamp, dest): plt.savefig(plotname, bbox_inches="tight") def load_data(self, wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List: - def convert_to_wav(input_path, output_path): - from pydub import AudioSegment - try: - audio = AudioSegment.from_mp3(input_path) - audio.export(output_path, format="wav") - print("音频文件为mp3格式,已转换为wav格式") - - except Exception as e: - print(f"转换失败:{e}") - def load_wav(path: str) -> np.ndarray: - if not path.lower().endswith('.wav'): - import os - input_path = path - path = os.path.splitext(path)[0]+'.wav' - convert_to_wav(input_path,path) #将mp3格式转换成wav格式 - waveform, _ = librosa.load(path, sr=fs) return waveform diff --git a/runtime/python/onnxruntime/funasr_onnx/paraformer_online_bin.py b/runtime/python/onnxruntime/funasr_onnx/paraformer_online_bin.py index 6c86bcd464..4d353b7336 100644 --- a/runtime/python/onnxruntime/funasr_onnx/paraformer_online_bin.py +++ b/runtime/python/onnxruntime/funasr_onnx/paraformer_online_bin.py @@ -219,24 +219,7 @@ def infer(self, feats: np.ndarray, feats_len: np.ndarray, cache): return asr_res def load_data(self, wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List: - - def convert_to_wav(input_path, output_path): - from pydub import AudioSegment - try: - audio = AudioSegment.from_mp3(input_path) - audio.export(output_path, format="wav") - print("音频文件为mp3格式,已转换为wav格式") - - except Exception as e: - print(f"转换失败:{e}") - def load_wav(path: str) -> np.ndarray: - if not path.lower().endswith('.wav'): - import os - input_path = path - path = os.path.splitext(path)[0]+'.wav' - convert_to_wav(input_path,path) #将mp3格式转换成wav格式 - waveform, _ = librosa.load(path, sr=fs) return waveform diff --git a/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py b/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py index 0bc265ebb5..d3630db5ac 100644 --- a/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py +++ b/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py @@ -202,24 +202,7 @@ def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs): return asr_res def load_data(self, wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List: - - def convert_to_wav(input_path, output_path): - from pydub import AudioSegment - try: - audio = AudioSegment.from_mp3(input_path) - audio.export(output_path, format="wav") - print("音频文件为mp3格式,已转换为wav格式") - - except Exception as e: - print(f"转换失败:{e}") - def load_wav(path: str) -> np.ndarray: - if not path.lower().endswith('.wav'): - import os - input_path = path - path = os.path.splitext(path)[0]+'.wav' - convert_to_wav(input_path,path) #将mp3格式转换成wav格式 - waveform, _ = librosa.load(path, sr=fs) return waveform diff --git a/runtime/python/onnxruntime/funasr_onnx/vad_bin.py b/runtime/python/onnxruntime/funasr_onnx/vad_bin.py index 5d9d8fd4e2..74cfa2944a 100644 --- a/runtime/python/onnxruntime/funasr_onnx/vad_bin.py +++ b/runtime/python/onnxruntime/funasr_onnx/vad_bin.py @@ -150,24 +150,7 @@ def __call__(self, audio_in: Union[str, np.ndarray, List[str]], **kwargs) -> Lis return segments def load_data(self, wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List: - - def convert_to_wav(input_path, output_path): - from pydub import AudioSegment - try: - audio = AudioSegment.from_mp3(input_path) - audio.export(output_path, format="wav") - print("音频文件为mp3格式,已转换为wav格式") - - except Exception as e: - print(f"转换失败:{e}") - def load_wav(path: str) -> np.ndarray: - if not path.lower().endswith('.wav'): - import os - input_path = path - path = os.path.splitext(path)[0]+'.wav' - convert_to_wav(input_path,path) #将mp3格式转换成wav格式 - waveform, _ = librosa.load(path, sr=fs) return waveform diff --git a/tests/test_runtime_eval_native.py b/tests/test_runtime_eval_native.py index 62747fb2c1..e0e6154800 100644 --- a/tests/test_runtime_eval_native.py +++ b/tests/test_runtime_eval_native.py @@ -273,3 +273,40 @@ def test_hotword_wrappers_decode_every_batch( assert [row["preds"][0] for row in alone] == ["你" * 4, "你" * 11, "你" * 17] # More files than batch_size: every batch must return its results. assert model(paths, hotwords) == alone + + +@pytest.mark.parametrize("extension", ["flac", "ogg", "mp3"]) +def test_onnx_wrappers_read_non_wav_paths(native_fixture, tmp_path, monkeypatch, extension): + np = pytest.importorskip("numpy") + librosa = pytest.importorskip("librosa") + soundfile = pytest.importorskip("soundfile") + if extension.upper() not in soundfile.available_formats(): + pytest.skip(f"libsndfile {soundfile.__libsndfile_version__} cannot write {extension}") + directory, _, _ = native_fixture + monkeypatch.syspath_prepend(str(ROOT / "runtime/python/onnxruntime")) + funasr_onnx = importlib.import_module("funasr_onnx") + online = importlib.import_module("funasr_onnx.paraformer_online_bin") + audio = tmp_path / f"talk.{extension}" + tone = np.sin(np.arange(48000) * 2 * np.pi * 440 / 16000).astype(np.float32) * 0.1 + soundfile.write(str(audio), tone, 16000) + # A different recording with the same stem must not be read in its place. + soundfile.write(str(tmp_path / "talk.wav"), np.zeros(16000, dtype=np.float32), 16000) + files = {path.name: path.read_bytes() for path in tmp_path.iterdir()} + expected = librosa.load(str(audio), sr=16000)[0] + wrappers = [ + funasr_onnx.Paraformer, + funasr_onnx.ContextualParaformer, + funasr_onnx.SeacoParaformer, + online.Paraformer, + funasr_onnx.Fsmn_vad, + funasr_onnx.SenseVoiceSmall, + ] + for wrapper in wrappers: + # load_data uses no model state, so skip __init__ and its model files. + waveforms = wrapper.__new__(wrapper).load_data(str(audio), 16000) + assert len(waveforms) == 1 + name = f"{wrapper.__module__}.{wrapper.__name__}" + np.testing.assert_array_equal(waveforms[0], expected, err_msg=name) + 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