实时ASR识别

[907122187]

在使用实时ASR识别过程中，发现deepspeech2online_aishell的识别效果很差，基本无法识别。使用conformer_online_multicn模型的过程中，感觉效果也比较一般。之前测试非实时ASR，觉得conformer_wenetspeech模型效果很好，实时识别的效果，无法达到conformer_wenetspeech的效果，请问有什么方法可以优化？

以下是实时deepspeech2online_aishell模型对官网提供的例子进行识别的服务器输出：

[2022-04-22 13:25:37,115] [ INFO] - start to do model forward
[2022-04-22 13:25:37,115] [ INFO] - start to decoce one chunk with deepspeech2 model
[2022-04-22 13:25:37,126] [ INFO] - decode one best result: 国人无论跑步最重要的就是给我们带来了身体健康
[2022-04-22 13:25:37,126] [ INFO] - start to decoce one chunk with deepspeech2 model
[2022-04-22 13:25:37,136] [ INFO] - decode one best result: 国人无论跑步最重要的就是给我们带来了身体健康
[2022-04-22 13:25:37,157] [ INFO] - The connection remain the audio samples: (1440,)
[2022-04-22 13:25:37,158] [ INFO] - process the audio feature success, the connection feat shape: [1, 11, 161]
[2022-04-22 13:25:37,158] [ INFO] - After extract feat, the connection remain the audio samples: (160,)
[2022-04-22 13:25:37,158] [ INFO] - Required decoding window 7 frames, and the connection has 11 frames
[2022-04-22 13:25:37,158] [ INFO] - start to do model forward
[2022-04-22 13:25:37,158] [ INFO] - start to decoce one chunk with deepspeech2 model
[2022-04-22 13:25:37,169] [ INFO] - decode one best result: 国人无论跑步最重要的就是给我们带来了身体健康
[2022-04-22 13:25:37,169] [ INFO] - start to decoce one chunk with deepspeech2 model
[2022-04-22 13:25:37,180] [ INFO] - decode one best result: 国人无论跑步最重要的就是给我们带来了身体健康
[2022-04-22 13:25:37,187] [ INFO] - Required decoding window 7 frames, and the connection has 3 frames
[2022-04-22 13:25:37,187] [ INFO] - flast {num_frames} is less than context {context} frames, and we cannot do model forward

[Jackwaterveg]

流式的 conformer_wenetspeech 模型正在训练中，如果效果好的话，会公布出来。 conformer_online_multicn 和 deepspeech2online_aishell 使用的训练数据集没有 conformer_wenetspeech 大， 所以鲁棒性弱一些，建议在安静环境下使用。