SETUP-Cherokee.md

Setup

Example steps to run this project

Get the code

cd ~/git
git clone git@github.com:CherokeeLanguage/Multilingual_Text_to_Speech.git

Virtual Environment

cd ~/git/Multilingual_Text_to_Speech
conda create --prefix ./env python=3.7
conda activate ./env
pip install -r requirements.txt

Datasets

cd ~/git/Multilingual_Text_to_Speech
cd data/css10

cut -f 2 -d '|' val.txt | sort -u | while read l; do
    mkdir "$l"
done

Manually unpack zips into correct folders

The unpacked audio files take up about 86GB of space unpacked before spectrograms and other processing is performed.

Verify all needed audio files are in place

cd ~/git/Multilingual_Text_to_Speech
cd data/css_comvoi
cut -f 4 val.txt | while read f; do
    if [ ! -f "$f" ]; then echo "FILE MISSING: $f"; exit -1; fi
done

cd ~/git/Multilingual_Text_to_Speech
cd data/css_comvoi
cut -f 4 train.txt | while read f; do
    if [ ! -f "$f" ]; then echo "FILE MISSING: $f"; exit -1; fi
done

Patch phonemizer

For better fidelity in phonemic transcriptions we need to ensure the phonemizer call has with_stress set to "True".

--- a/utils/text.py
+++ b/utils/text.py
@@ -88,7 +88,7 @@ def to_phoneme(text, ignore_punctuation, language, phoneme_dictionary=None):
 def _phonemize(text, language):
     try:
         seperators = Separator(word=' ', phone='')
-        phonemes = phonemize(text, separator=seperators, backend='espeak', language=language)           
+        phonemes = phonemize(text, separator=seperators, backend='espeak', with_stress=True, language=language)           
     except RuntimeError:
         epi = epitran.Epitran(language)
         phonemes = epi.transliterate(text, normpunc=True)

Prepare spectrograms:

cd ~/git/Multilingual_Text_to_Speech
conda activate ./env
cd data
python3 prepare_css_spectrograms.py

Adjust hyper params

cd ~/git/Multilingual_Text_to_Speech
conda activate ./env #only if env is not already active
cd params
cp generated_training.json generated_training_low_memory.json 
code generated_training_low_memory.json 

-	"batch_size": 60,
+	"batch_size": 10,

-	"version": "GENERATED-TRAINING"
+	"version": "GENERATED-TRAINING-LOW-MEMORY"

Training

cd ~/git/Multilingual_Text_to_Speech
conda activate ./env #only if env is not already active

export PYTHONIOENCODING=utf-8
python train.py --hyper_parameters generated_training_low_memory