Speech Synthesis

• Speech synthesis- the use of a machine, usually a computer, to produce human-like speech

• Canned speech- prerecorded utterances and phrases

• Synthesized speech- piecing together smaller recorded units of speech into new utterances

• Intelligibility- how well listeners can recognize and understand the individual sounds or words generated by the synthesis system

• Naturalness- how much the synthesized speech sounds like the speech of an actual person

• Articulatory synthesis- a synthesis technique that generates speech “from scratch” based on computational models of the shape of the human vocal tract and the articulation processes

• Source-filter theory- there are two independent parts to the production of speech sounds

  • Source- the mechanism that creates a basic sound

  • Filter- shapes the sound created by the source into the different sounds we recognize as speech sounds

• Concatenative Synthesis- uses recorded speech by stringing together pieces of the recorded speech and then smoothing the boundaries between them

  • Unit selection synthesis- takes large samples of speech and builds a database of smaller units from these speech samples

  • Diphone synthesis- pairs of adjacent sounds are attached at the end of one phone and the beginning of another

  • Domain-specific synthesis- create utterances from prerecorded words and phrases that closely match the words and phrases that will be synthesized

• Text-to-speech synthesis- speech generated directly from text entered with normal orthography

Automatic Speech Recognition

• Automatic Speech Recognition- the conversion of an acoustic speech waveform into text

• Noisy channel model- treats speech input as if it has been passed through a communication channel that garbles the speech waveform

• Components of an Automatic Speech Recognition System:

  • Signal processing- recording the speech waveform with a microphone and storing it in a manner that is suitable for further processing by a computer

  • Acoustic modeling- mapping the energy values extracted during signal processing

  • Pronunciation modeling- used to filter out unlikely sound sequences

  • Language modeling- calculating the probability of sequences

• Parameters of Speech Recognition Systems

  • Speaking mode- only accepts isolated word input or continuous speech input

  • Vocabulary size- the size of the system’s vocabulary will impact its accuracy

  • Speaker Enrollment- the system may or may not need to be trained to a specific voice

Communicating with Computers

• Interactive Text-Bases Systems- dialogue carried between computer and user via text

  • Word spotting- a program focuses on words it knows and ignores ones it doesn’t

• Spoken-Language Dialogue Systems- Dialogue carried between computer and user via speech

  • Isolated speech- the user speaks the input clearly and without extraneous words

  • Continuous speech- input can be more like normal speech

• Components of a Spoken-language Dialogue System

  • Automatic Speech Recognition- combining levels of linguistic knowledge in order to allow speaker-independent understanding of continuous speech

  • Language Processing and Understanding- the system must decipher not only individual words, but also the intention of the speaker

  • Dialogue Management- the system needs to understand the intentional structure of the conversation

  • Text Generation- the use of computers to respond to humans using natural language by creating sentences that convey the relevant information

  • Speech Synthesis- the words that make up the generated text must be converted into a sequence of sounds

Machine Translation

• Translation- the task of converting the contexts of a text written in one language into a text in another language

• Machine translation- the use of computers to carry our translation

• Problems:

  • Context can often be removed

  • Lexical ambiguity

• Partial Automation- the source language text can first be pre-edited by a person so as to “prime” it for a machine translation system

Corpus Linguistics

• Corpus- a collected body of text

• Corpus linguistics- involves the design and the annotation of corpus materials that are required for specific purposes

• Corpus can be composed from spoken, signed or written language

• Corpora can be classified by the genre of the source material

• Balanced corpora- corpora that try to remain balanced among different genres

• Reference corpus- specified amount of text that has been collected and annotated

• Monitor corpus- as new texts continue to be written or spoken, more data is gathered