speechT

speetcht

用TensorFlow编写的OpenSource语音到文本软件。在Librispeech测试语料库中，达到8％的字母错误率，单词错误率为20％ 。

安装

先决条件

需要Python3，Portaudio19-Dev和FFMPEG。

在Ubuntu安装上

 sudo apt install python3-pip portaudio19-dev ffmpeg

通过PIP3安装

 pip3 install git+https://github.com/timediv/speechT

建筑学

当前的SpeechT基于WAV2LETTER纸张和CTC损耗函数。

来自http://www.opensl.org/12/的语音语料库自动下载。
注意：语料库约为30GB！

训练

数据必须在培训之前进行预处理

 speecht-cli preprocess

然后，进行培训，执行

 speecht-cli train

使用--help以获取更多详细信息。

您可以监视培训并在张板中查看其他日志

 tensorboard --logdir log/

测试

在整个测试集运行中评估

 speecht-cli evaluate

评估一批

 speecht-cli evaluate --step-count 1

默认情况下，使用贪婪解码。请参阅有关如何使用Kenlm进行解码的Using a language model部分。

使用--help以获取更多详细信息。

实时用法

使用麦克风记录并打印预测运行

 speecht-cli record

使用--help以获取更多详细信息。

训练有素的重量

您没有自己的资源可以自己训练？从这里下载权重

 mkdir train
tar xf speechT-weights.tgz -C train/

然后，您可以将模型与EG评估一起使用

 speecht-cli evaluate --run-name best_run

使用语言模型

如果您想将KENLM用作解码的语言模型，则需要编译并安装tensorflow-with-kenlm。如果您仅需要Linux的CPU版本的TensorFlow版本，则也可以在此处下载。

从这里下载所有必要的文件，然后

 tar xf kenlm-english.tgz
speecht-cli evaluate --language-model kenlm-english/

结果

默认参数在NVIDIA TITAN X上进行了约5至6天的训练。

总体统计数据

 Average Letter Edit Distance: 7.7125
Average Letter Error Rate: 8%
Average Word Edit Distance: 3.801953125
Average Word Error Rate: 20%

在一些示例中进行的预测和预测

 expected: but that is kaffar's knife
decoded: but that is caffr's klife 
LED: 4 LER: 0.15 WED: 2 WER: 0.40

expected: he moved uneasily and his chair creaked
decoded: he moved uneasily in his chair creet
LED: 5 LER: 0.13 WED: 2 WER: 0.29

expected: it is indeed true that the importance of tact and skill in the training of the young and of cultivating their reason and securing their affection can not be overrated
decoded: it is indeed true that the importance of tact and skill in the training of the young and of cultivating their reason and so carrying their affection can not be o rated
LED: 8 LER: 0.05 WED: 4 WER: 0.13

expected: she pressed his hand gently in gratitude
decoded: she pressed his hand gently in gratitude
LED: 0 LER: 0.00 WED: 0 WER: 0.00

expected: don't worry sizzle dear it'll all come right pretty soon
decoded: don't worry i l dear it all come riprety soon 
LED: 13 LER: 0.23 WED: 5 WER: 0.50

expected: may we see gates at once asked kenneth
decoded: may we see gates at once asked keneth 
LED: 2 LER: 0.05 WED: 1 WER: 0.12

可以在此处找到整个评估日志。