已经训练好的HMM模型用于单词识别

Question

我使用自动编码器实现了一个音素分类器（给定一个音频文件数组，它返回所有已识别的音素）。我想扩展这个项目，以便识别单词。是否存在已经训练过的HMM模型（英文），它会识别给出音素列表的单词？

谢谢大家。

Answer 1

I am not aware of any decoder that could help you. Speech recognition software does not work this way.

Usually such thing requires custom implementation for dynamic beam search. That is not a huge task, maybe 100 lines of code. It also depends on what your phonetic decoder produces. Is it phonetic lattice (ideally) or is it a 1-best result with scores or simply 1-best result without scores.

In case you have a proper lattice you might want to try openfst toolkit where you convert LM and dictionary to FST, then compose with lattice FST and then use fstbestpath to find the best path. Still, instead of all those phonetic conversions you can simply write a dynamic search.

Baidu in their projects also convert speech to letters and then use language model to fix letter sequence. But they say that without langauge model it works equally well.