"Personalising synthetic voices for individuals with severe speech impairment"

Chapter 3

Example 3.1: Articulatory synthesis (Taken from D. Hill, L. Manzara and C. Schock (1995) "Real-time articulatory speech-synthesis-by-rules", In Proceedings of AVIOS 1995, San Jose, CA: USA, September 1995, pp27-44, via http://pages.cpsc.ucalgary.ca/~hill/papers/avios95/index.htm)

Example 3.2: The JSRU/Holmes synthesiser (Taken from S. Lemmety (1999) "Review of speech synthesis technology" Master's thesis, Helsinki University of Technology, via http://www.acoustics.hut.fi/publications/files/theses/lemmetty_mst/appa.html

Example 3.3: The KlatTalk system, introduced by D. Klatt (Taken from D. Klatt (1987), "Review of text-to-speech conversion for English" Journal of the Acoustical Society of America 82, 737-793, via http://festvox.org/history/klatt.html)

Example 3.4: Several DECTalk voices, introduced by D. Klatt (Taken from D. Klatt (1987), "Review of text-to-speech conversion for English" Journal of the Acoustical Society of America 82, 737-793, via http://festvox.org/history/klatt.html)

Example 3.5: Limited domain synthesis from Festival (Taken from http://festvox.org/ldom/ldom_time.html)

Example 3.6: General Festival concatenative synthesis (Taken from http://www.cs.cmu.edu/~awb/festival_demos/general.html)

Example 3.7: The author's original speech for comparison with examples 3.8, 3.9 and 3.13.

Example 3.8: Festvox voice built with Arctic set A (593 sentences/80 minutes) of the author's recorded data.

Example 3.9: ModelTalker voice built with approximately 1800 utterances (40-50 minutes) of the author's speech.

Example 3.10: Speaker 1's original speech for comparison with examples 3.11 and 3.12.

Example 3.11: Festvox voice built with Arctic set A (593 sentences/80 minutes) of speaker 1's data, showing where concatenative synthesis can break down. Please listen to the sound file then click on the footnote to see what the voice was saying^[1]. A different example of this speaker's voice is found as example 3.11.

Example 3.12: Festvox voice built with Arctic set A (593 sentences/80 minutes) of speaker 1's data.

Example 3.13: HTS voice built with Arctic set A (593 sentences/80 minutes) of the author's data.

Chapter 5

Example 5.1: Speaker 1's original speech.

Example 5.2: Speaker 2's original speech.

Example 5.3: Average voice from which the models were adapted towards the target speakers.

Example 5.4: Experiment stimuli using 10 sentences of adaptation data taken from speaker 1.

Example 5.5: Experiment stimuli using 100 sentences of adaptation data taken from speaker 1.

Example 5.6: Experiment stimuli using 500 sentences of adaptation data taken from speaker 1.

Example 5.7: Experiment stimuli resynthesised original target speech from speaker 1.

Example 5.8: Average voice from which the models were adapted towards the target speakers.

Example 5.9: Experiment stimuli using 10 sentences of adaptation data taken from speaker 2.

Example 5.10: Experiment stimuli using 100 sentences of adaptation data taken from speaker 2.

Example 5.11: Experiment stimuli using 500 sentences of adaptation data taken from speaker 2.

Example 5.12: Experiment stimuli resynthesised original target speech from speaker 2.

Chapter 6

Example 6.1: Original speech taken from speaker 3. Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[3].

Example 6.2: Original speech taken from speaker 4. Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[4].

Example 6.3: Original speech taken from speaker 5. Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[5].

Example 6.4: Speaker 5's synthesised voice built with unedited data and using all his own voice features. Please listen to this sound file and example 6.5 then click on the footnote if you are having difficulty understanding the speech^[2].

Example 6.5: Speaker 5's synthesised voice built with edited data and using all his own voice features. Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[2].

Example 6.6: Example of the Acapela voice, Peter, with which the speaker participants compared their own synthesised voices.

Example 6.7: Voice built with average voice features and speaker 3's own spectral features and log F0, using edited data. Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[6].

Example 6.8: Voice built with average voice features and speaker 4's own spectral features and log F0, using edited data. Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[7].

Example 6.9: Voice built with average voice features and speaker 5's own spectral features and log F0, using edited data. Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[8].

Example 6.10: Voice built with average voice features and speaker 5's own spectral features and log F0, using unedited data (for comparison with example 6.9). Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[8].

Example 6.11: Voice built with average voice features and speaker 3's own spectral features, log F0 and energy (for comparison with example 6.7). Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[6].

Example 6.12: Voice built with average voice features and speaker 4's own spectral features, log F0 and energy (for comparison with example 6.8). Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[7].

Example 6.13: Voice built with average voice features and speaker 3's own spectral features, log F0 and global variance for log F0 (for comparison with example 6.7). Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[6].

Example 6.14: Voice built with average voice features and speaker 4's own spectral features, log F0 and global variance for log F0 (for comparison with example 6.8). Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[7].

Example 6.15: Voice built with average voice features and speaker 5's own spectral features, log F0 and global variance for log F0 (for comparison with example 6.9). Please listen to the sound file then click on the footnote if you are having difficulty understanding the speech^[8].

^[1] Text of sound file: "I feel like a kid on a long car journey playing I-spy in the fog."

^[2] Text of sound file: "Horses and rifles had been her toys, camp and trail her nursery."

^[3] Text of sound file: "The fourth and fifth days passed without any developments."

^[4] Text of sound file: "In spite of their absurdity the words affected Philip curiously."

^[5] Text of sound file: "Earth and gravel seemed to fill the pan."

^[6] Text of sound file: "The government triumphed four years ago and we have every reason to believe that it will triumph again."

^[7] Text of sound file: "We have proof that the regime wields sufficient power in the North to exploit the entire population."

^[8] Text of sound file: "She had scarcely divulged the scandal before it was splattered over the front pages of the tabloids."

"Personalising synthetic voices for individuals with severe speech impairment"

Sarah Creer PhD thesis - sound files

Chapter 3

Chapter 4

Chapter 5

Chapter 6