识别效果

[Aimeelulu]

识别效果

[yfq512]

@Aimeelulu 大神好，global_cmvn、final.pt、lang_char.txt这三个文件怎么生成啊？

[yfq512]

我用了202108……的那个模型识别效果还不错。

[phoenixsfly]

请问怎么改动，我用官方最新的训练模型会出现错误，是新版预训练模型对模型结构调整了导致无法加载吗？

RuntimeError: Error(s) in loading state_dict for ASRModel:
        Missing key(s) in state_dict: "encoder.encoders.0.conv_module.norm.running_mean", "encoder.encoders.0.conv_module.norm.running_var", "encoder.encoders.1.conv_module.norm.running_mean", "encoder.encoders.1.conv_module.norm.running_var", "encoder.encoders.2.conv_module.norm.running_mean", "encoder.encoders.2.conv_module.norm.running_var", "encoder.encoders.3.conv_module.norm.running_mean", "encoder.encoders.3.conv_module.norm.running_var", "encoder.encoders.4.conv_module.norm.running_mean", "encoder.encoders.4.conv_module.norm.running_var", "encoder.encoders.5.conv_module.norm.running_mean", "encoder.encoders.5.conv_module.norm.running_var", "encoder.encoders.6.conv_module.norm.running_mean", "encoder.encoders.6.conv_module.norm.running_var", "encoder.encoders.7.conv_module.norm.running_mean", "encoder.encoders.7.conv_module.norm.running_var", "encoder.encoders.8.conv_module.norm.running_mean", "encoder.encoders.8.conv_module.norm.running_var", "encoder.encoders.9.conv_module.norm.running_mean", "encoder.encoders.9.conv_module.norm.running_var", "encoder.encoders.10.conv_module.norm.running_mean", "encoder.encoders.10.conv_module.norm.running_var", "encoder.encoders.11.conv_module.norm.running_mean", "encoder.encoders.11.conv_module.norm.running_var".

[yfq512]

我没修改，就是找了个模型，改几个配置参数，就可以跑了

[chengming1108]

我用了202108……的那个模型识别效果还不错。

我发现模型参数读不进来 直接四个都替换掉吗

[chengming1108]

我没修改，就是找了个模型，改几个配置参数，就可以跑了

您好 wenet_decoder.py 那几个参数怎么修改。可以分享一下吗 后面的参数读不进来

[yfq512]

@chengming1108

加载模型和字典等

……
    ASR_model = WeNetDecoder("20210815_unified_conformer_exp/train.yaml")
    with open('./test5.wav','rb') as audiostream:
        for dataflow in iter(lambda:audiostream.read(4000),""):
            if len(dataflow) == 0:
                break
            sig = struct.unpack("%ih" % (len(dataflow) / 2), dataflow)
            data = np.array([float(val) for val in sig])
            ASR_model.detect(data)
            ASR_model.ctc_prefix_beam_search_purn()
            #SR_model.ctc_greedy_search_purn()
            # time.sleep(0.1)
    s = ASR_model.decoder_recoring()
    print('result', s)
……

20210815_unified_conformer_exp/train.yaml 修改配置

……
cmvn_file: 20210815_unified_conformer_exp/global_cmvn
checkpoint: 20210815_unified_conformer_exp/final.pt
dict: 20210815_unified_conformer_exp/words.txt
……

[chengming1108]

@chengming1108

加载模型和字典等

……
    ASR_model = WeNetDecoder("20210815_unified_conformer_exp/train.yaml")
    with open('./test5.wav','rb') as audiostream:
        for dataflow in iter(lambda:audiostream.read(4000),""):
            if len(dataflow) == 0:
                break
            sig = struct.unpack("%ih" % (len(dataflow) / 2), dataflow)
            data = np.array([float(val) for val in sig])
            ASR_model.detect(data)
            ASR_model.ctc_prefix_beam_search_purn()
            #SR_model.ctc_greedy_search_purn()
            # time.sleep(0.1)
    s = ASR_model.decoder_recoring()
    print('result', s)
……

20210815_unified_conformer_exp/train.yaml 修改配置

……
cmvn_file: 20210815_unified_conformer_exp/global_cmvn
checkpoint: 20210815_unified_conformer_exp/final.pt
dict: 20210815_unified_conformer_exp/words.txt
……

收到 谢谢