diff --git a/doc/design/speech/README.MD b/doc/design/speech/README.MD index 7304650e628dba210488cd2dc4836318b5383b2a..cc03aac7b4970a66a5c8c7591aecf6349ae92f8a 100644 --- a/doc/design/speech/README.MD +++ b/doc/design/speech/README.MD @@ -140,7 +140,17 @@ TODO by Assignees ### Beam Search with CTC and LM -TODO by Assignees +
+
+Figure 2. Algorithm for Beam Search Decoder. +
+ +- The **Beam Search Decoder** for DS2 CTC-trained network follows the similar approach in \[[3](#references)\] with a modification for the ambiguous part, as shown in Figure 2. +- An **external defined scorer** would be passed into the decoder to evaluate a candidate prefix during decoding whenever a space character appended. +- Such scorer is a unified class, may consisting of language model, word count or any customed evaluators. +- The **language model** is built from Task 5, with a parameter should be carefully tuned to achieve minimum WER/CER (c.f. Task 7) +- This decoder needs to perform with **high efficiency** for the convenience of parameters tuning and speech recognition in reality. + ## Future Work @@ -153,3 +163,4 @@ TODO by Assignees 1. Dario Amodei, etc., [Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin](http://proceedings.mlr.press/v48/amodei16.pdf). ICML 2016. 2. Dario Amodei, etc., [Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin](https://arxiv.org/abs/1512.02595). arXiv:1512.02595. +3. Awni Y. Hannun, etc. [First-Pass Large Vocabulary Continuous Speech Recognition using Bi-Directional Recurrent DNNs](https://arxiv.org/abs/1408.2873). arXiv:1408.2873 diff --git a/doc/design/speech/image/beam_search.png b/doc/design/speech/image/beam_search.png new file mode 100644 index 0000000000000000000000000000000000000000..7f7e35f34223162d0f7f0ed97375909c43b830ae Binary files /dev/null and b/doc/design/speech/image/beam_search.png differ