The Basic Model

Speech Recognition is split into two main phases: transforming the signal into phones, and then the phones into words.

Signal processing:

• analog speech signal (energy) to digital

• sampling rate of 8-16 KHz (8000-16000 times per second)

• The quantization factor determines the precision of the sample and is 8 to 12 bits

• 8000 samples/sec * 8 bits/sample = 64000 bits/sec = 8000 bytes/sec which yields ~0.5 MBytes / minute

• speaker environment greatly affects the performance
  • accents, vocal tracts

  • amount of background noise

  • filter out for general-purpose speech recognition

  • need for speaker identification

• Define each word's pronunciation as a sequence of phones.

• Then transforming phones to words can be done by lookup.

• Homophones are two words with same sound, e.g. heh, hay. There are also words with multiple pronunciations, e.g. Caribbean.

• Segmentation is separation between words. A fluent language has little silence.