New improved recipes

There are several recipes which build on and improve the naive one already described. Some examples are given here.

naive_glott: naive recipe with GlottHMM vocoder

## Assuming that you want to start from scratch:
rm -r ./train/rm/speakers/rss_toy_demo/naive_glott/ ./voices/rm/rss_toy_demo/naive_glott/

## Train:
python ./scripts/train.py -s rss_toy_demo -l rm -text wikipedia_10K_words naive_glott

## Synthesise:
./scripts/speak.py -l rm -s rss_toy_demo -o ./test/wav/romanian_toy_naive_glott.wav \
                                        -play naive_glott ./test/txt/romanian.txt

This is the same as the naive recipe but uses the high-quality vocoder GlottHMM for speech analysis and synthesis.

naive_glott_prom: wavelet-based prominence labelling

## Assuming that you want to start from scratch:
rm -r ./train/rm/speakers/rss_toy_demo/naive_glott_prom/ ./voices/rm/rss_toy_demo/naive_glott_prom/

## Train:
python ./scripts/train.py -s rss_toy_demo -l rm -text wikipedia_10K_words naive_glott_prom

## Synthesise:
./scripts/speak.py -l rm -s rss_toy_demo -o ./test/wav/romanian_toy_naive_glott_prom.wav \
                                        -play naive_glott_prom ./test/txt/romanian.txt

This is the same as the naive_glott recipe but also makes use of an unsupervised representation of prominence similar to the one described here. Extraction of the representation is based on wavelet transform-derived acoustic features and prediction makes use of vector space models of words and a decision tree classifier.

Voices from non-alphabetic script data

A Hindi toy corpus (extracted from the IIIT Indic database available here) is included to demonstrate parts of the recipe developed for the Simple4All Blizzard Challenge entry described in this paper. The recipes blizzard_2014_naive_latinised and blizzard_2014_naive_latinised_syl incrementally introduce the naive alphabetisation and syllabification described in the paper. Due to the toy corpus’s small size, the syllabification severely affects the quality of the speech. The recipe blizzard_2014_naive_latinised_glott adds the latinisation and GlottHMM vocoder:

## Assuming that you want to start from scratch:
rm -r ./train/hi/speakers/toy/blizzard_2014_naive_latinised_glott/ ./voices/hi/toy/blizzard_2014_naive_latinised_glott/

## Train:
python ./scripts/train.py -s toy -l hi -text wikipedia_10K_words blizzard_2014_naive_latinised_glott

## Synthesise:
./scripts/speak.py -l hi -s toy -o ./test/wav/hindi_naive_latinised_glott.wav \
             -play blizzard_2014_naive_latinised_glott ./test/txt/hindi.txt

A simpler recipe like the naive one can be used here for comparison:

## Assuming that you want to start from scratch:
rm -r ./train/hi/speakers/toy/naive/ ./voices/hi/toy/naive/

python ./scripts/train.py -s toy -l hi -text wikipedia_10K_words naive

 ./scripts/speak.py -l hi -s toy -o ./test/wav/hindi_naive.wav \
                -play naive ./test/txt/hindi.txt