Bounty hunter

Thursday, 5 October 2017


During the creation of our alpha version, we made a series of comparison tests. These were the results.
Speech recognition systems used for testing:
  • MS Azure
  • Google Speech
  • IBM Watson
  • Anryze STT V.b0.53
For testing, the following variants were applied:
  • 3 male voices
  • 2 female voices
  • Separate speech (pause after every word)
  • Continuous speech
  • General topics
  • Special word usage (e.g. international names, scientific and technical terms, etc.)
Tests description:
  1. 1 speaker, continuous speech, general topic
  2. 1 speaker, separate speech, general topic
  3. 1 speaker continuous speech, special words
  4. 1 speaker, separate speech, special words
  5. 2 speakers, continuous speech, general topic
  6. 2 speakers, separate speech, general topic
  7. 2 speakers, continuous speech, general topic
  8. 2 speakers, separate speech, special words
Before the start of testing we prepared 10 different text sets for each variant.
Testing process
Each speaker read out particular texts according to the test type. Recordings were made in a recording studio in order to ensure clean sound. After recording, the sound files were converted into appropriate formats for each recognition system. The results are displayed in the chart below. For the second stage of testing we added background noise. The results of testing in a noisy environment are shown in the second chart.

No comments:

Post a Comment