Menu

Pocketsphinx Decoding Time with unigram, bigram and trigram models

Help
2015-11-13
2015-11-13
  • Van Nhan Nguyen

    Van Nhan Nguyen - 2015-11-13

    Hi,

    I trained unigram, bigram and trigram language models and tested them with the Pocketsphinx recognizer. The results were quite good, however, the decoding time of the models were weird.

    Average Decoding Time and RTF
    Unigram: 3.33s (RFT: 0.87)
    Bigram: 2.71s (RTF:0.74)
    Trigram: 2.58s (RTF:0.70)

    Is the Unigram model supposed to be the fastest among the three models?

    Best Regards.
    Nhan.

     
    • Nickolay V. Shmyrev

      Is the Unigram model supposed to be the fastest among the three models?

      It is not that simple.

      Pocketsphinx uses complex speech recognition algorithm with different kinds of pruning and interoperation between multi-pass searches. So unigram model with less accuracy might require more time just because it can not prune hypothesis effectively.

       

Log in to post a comment.