Chapter 7.2 Speech Synthesis

Front 1

What is "canned speech" and when it can be used?

Back 1

The speech signals recorded and saved as they are. Playback with as they are, segment boundaries can be manipulated a little. Used when vocabulary small.

Front 2

What are the classes of speech output systems? Describe them. What do they convert to speech? What steps do they include?

Back 2

1. Announcement machine: phonetics to speech
2. Statement machine: concept-to-speech (CTS). Avoids some fof the problems of reading out machine, thus a more natural speech achieved.
3. Reading out machine: text to speech (TTS). The simplest case. Symbolic processing, generates prosody and generates speech signal.

Front 3

What steps does symbolic processing have?

Back 3

1. Preprocessing: write out abbreviation, distinguish main clauses and subordinate clauses
2. Exception handling: names, foreign-language words > other pronunciation ("multilingual synthesis").
3. Morphological analysis: better sentence stress
4. Assigning word stress: rules or lexicon based stress finding
5. Orthography-to-phonetic mapping
6. Word classes identified
7. Syntactic and prosodic structure analysed. Prosody applied based on punctuation, word stress, syntactic objects

Front 4

What are the two approaches for speech prosody generation?

Back 4

1. Based on rules
- Fujisaki: Word and phrase components superimposed on a declination line, processed by a 2nd-order system
- Adriaens: Copy contours
- Mersdorf: LPC parametrization
2. Based on data
- Neural networks
- Classification trees

Front 5

Explain how formant synthesis works. What are the control parameters?

Back 5

Idealized excitation signals: voiced impulse comb and unvoiced noise.

Two formant filter lines: longer (3-5 filters/resonators) for vowels, shorter (1-2 filters/resonators) for fricatives. Aspirated sounds use both.

Control parameters: fundamental frequency, frequencies and bandwidth of formant filters, amplitudes of excitation signals.

Front 6

How does LPC synthesizer work? What are its control parameters?

Back 6

Front 7

What is articulatory synthesis?

Back 7

Parametric synthesis, modelling the exact movement of articulators in vocal tract.

Computationally heavy. Source-filter model picture (diameters of different parts in vocal tract).

Front 8

What are advantages and disadvantages of parametric syntesis? Name another type of syntesis.

Back 8

1. Formant synthesis: Difficult determination of parameter values
2. LPC Synthesis: Only 2 types of excitation -> small variability, but simple generation
3. Articulatory Synthesis: Detailed modelling of the control and the movements of the human vocal tract

Another type: Concatenative Synthesis - concatenation of individual elements (phones, diphones, demisyllables, syllables, etc.)

Front 9

What is concatenation synthesis? How does it work?

Back 9

Lots of pre-recorded material, choosing short pieces and concatenating so that the transition are smooth.

Front 10

How does PSOLA work? What are the benefits of PSOLA in speech synthesis?

Back 10

Recorded signal cut into elementary components (at fundamental period markers), which are overlaid and added to make a new signal with different fundamental frequency. Fundamental frequency and thus prosody can be adjusted by elementary component distances. Smooth changes of f0 makes the speech more natural. However, manipulation causes artefacts.

Front 11

What is unit-selection synthesis?

How long are the units?

How does it decide what units to select?

What kind of labelling is needed?

When does it work correctly and when not?

Back 11

Choose as long speech units as possible without manipulating.

Requires lots of recordings (high effort).

Units that 1) match the text to be synthesized 2) can be used together with minimal discontinuities in signal are chosen, based on a cost function.

The costs are called 1) cost of units 2) concatenation costs. Units need to be labelled both phonemically and prosodically.

Text to be synthesized needs to have labels too, in real-time. Unit-selection achieves quite natural speech if the text is found in the inventory.

Front 12

How can HMM be used for speech synthesis?

Back 12

1. Best-fitting connection of elements/units
->Representation as an HMM
-> Elements = States of the HMM, each linked to a parametric signal representation (e.g. LPC)
2. Synthesis = Finding the optimum path through all states (elements)