The lmDecode action decodes recurrent neural network (RNN) scores using a specified language model. It's a crucial step in speech-to-text systems, refining the raw output of an acoustic model by applying linguistic rules and probabilities to generate more coherent and accurate text.
| Parameter | Description |
|---|---|
| alpha | Specifies a tunable parameter that controls the trade-off between the acoustic model and the language model. The larger 'alpha' is, the more the language model will count while ranking all candidates of the final guess. |
| beta | Specifies a tunable parameter that controls the trade-off between the acoustic model and the length of the sentence. The larger 'beta' is, the more the length of the sentence will count while ranking all candidates of the final guess. |
| blankLabel | Specifies the string used to indicate a 'blank' label. |
| casOut | Specifies the output table to store the decoded results. |
| columnMap | Specifies the labels that the score columns in each time frame represent, which must follow the same order. |
| copyVars | Specifies the variables to transfer from the input table to the output table. |
| langModelTable | Specifies the n-gram language model to use in the decoding process. |
| maxPathSize | Specifies the maximum number of paths kept as candidates for final results during the decoding process. |
| ngramsOrder | Specifies the highest order of n-grams to use during the decoding process. |
| spaceLabel | Specifies the string used to indicate a 'space' label. |
| table | Specifies the input table that contains the RNN scores returned by the acoustic model. |