请问为什么测试模型的result是空白的

[ccyniubi]

[MarStarck]

我也遇到这个问题了

[Mddct]

把flac 转成wav试下

[caiyuxi]

我也遇到类似的问题。

import torchaudio
import wenet
waveform, sample_rate = torchaudio.load("examples_1995-1836-0002.flac")
torchaudio.save(
    "sound2.wav", waveform, sample_rate,
    encoding="PCM_S")
model = wenet.load_model("english")
result = model.transcribe("sound2.wav")
print(result)

stdout为 {'text': '▁UM', 'confidence': 0.6211525614053677}
音频文件来自https://huggingface.co/spaces/wenet/wenet_demo/tree/main/examples
除了wenet.load_model("english")，其他的预训练模型也有类似的问题。

@pytest.mark.parametrize("model", [
    "gigaspeech_u2pp_conformer_libtorch.tar.gz",
    "librispeech_u2pp_conformer_libtorch.tar.gz",
])
def test_model(model):
    dest = model.split('.')[0]  # aishell_u2pp_conformer_libtorch
    dataset = model.split('_')[0]  # aishell
    if not os.path.exists(dest):
        os.makedirs(dest)
    response = requests.get(
        "https://modelscope.cn/api/v1/datasets/wenet/wenet_pretrained_models/oss/tree"  # noqa
    )
    model_info = next(data for data in response.json()["Data"]
                      if data["Key"] == model)
    model_url = model_info['Url']
    download(model_url, dest=dest, only_child=True)
    model = Model(dest, gpu=-1, beam=5, resample_rate=16000)
    assert dataset in ['gigaspeech', 'librispeech']
    audio_file = "sound2.wav"
    result = model.transcribe(audio_file)
    print(result)

gigaspeech_u2pp_conformer_libtorch.tar.gz": {'confidence': 0.6211525614053677, 'text': '▁UM'}
"librispeech_u2pp_conformer_libtorch.tar.gz": {'confidence': 0.5829262466384643, 'text': ''}