[Home]

Pass Rate Tests for CVC-contained and CVC-free Words (Figure 3)

Environment

• Identify 456 unique words from the latest list of Amazon Alexa commands [link]
• Identify 441 unique words from the latest list of Google Assistnat commands [link]
• Google TTS [link]
• Amazon Alexa [link]
• Google STT [link]

Experiment Procedure

• Step 1. Randomly select 50 CVC-contained and 50 CVC-free words
• Step 2. Generate audio files at different playback speed (1.0 – 3.0x) using Google TTS (Format and naming of each audio file: [word]_waveNetM_[speed].mp3, e.g., everyone_waveNetM_2.0.mp3)
  • CVC-contained
  • CVC-free
• Step 3. Transmit corresponding audio files over-the-wire to target ASR
  (i.e., Amazon Alexa or Google STT)
• Step 4. Compute pass rate = number of the non-empty ASR recognition / total number of tested words

Codes Utilized for Experiment

• Translate words into their phonetic representation
• CVC_CV_w2p_analyze.py
• Alexa word list
• Google word list
• Identify CVC-contained and CVC-free words via phaoneme syllabification
• Phoneme_analyzer.py
• CMU pronunciation dictionary

Result

• Figure 3 in the paper presents the pass rates for CVC-contained and CVC-free words