tacotron-cli 0.0.3

Command-line interface (CLI) to train Tacotron 2 using .wav <=> .TextGrid pairs.

tacotron-cli

Command-line interface (CLI) to train Tacotron 2 using .wav <=> .TextGrid pairs.

Features

• train phoneme stress separately (ARPAbet/IPA)
• train phoneme tone separately (IPA)
• train phoneme duration separately (IPA)
• train single/multi-speaker
• train/synthesize on CPU or GPU
• synthesis of paragraphs
• copy embeddings from one checkpoint to another
• train using embeddings or one-hot encodings

Installation

pip install tacotron-cli --user

Usage

usage: tacotron-cli [-h] [-v] {train,continue-train,validate,synthesize,analyze,add-missing-symbols} ...

Command-line interface (CLI) to train Tacotron 2 using .wav <=> .TextGrid pairs.

positional arguments:
  {train,continue-train,validate,synthesize,analyze,add-missing-symbols}
                              description
    train                     start training
    continue-train            continue training from a checkpoint
    validate                  validate checkpoint(s)
    synthesize                synthesize lines from a file
    analyze                   analyze checkpoint
    add-missing-symbols       copy missing symbols from one checkpoint to another

optional arguments:
  -h, --help                  show this help message and exit
  -v, --version               show program's version number and exit

Dependencies

• torch
• pandas
• numpy
• librosa
• plotly
• matplotlib
• scikit-image
• scikit-learn
• scipy
• tqdm
• ordered_set>=4.1.0 speech-dataset-parser>=0.0.1
• mel-cepstral-distance>=0.0.1

Training

The dataset structure need to follow the generic format of speech-dataset-parser, i.e., each TextGrid need to contain a tier in which all phonemes are separated into single intervals, e.g., T|h|i|s| |i|s| |a| |t|e|x|t|..

Tips:

• place stress directly to the vowel of the syllable, e.g. b|ˈo|d|i instead of ˈb|o|d|i (body)
• place tone directly to the vowel of the syllable, e.g. ʈʂʰ|w|a˥˩|n instead of ʈʂʰ|w|a|n˥˩ (串)
  • tone-characters which are considered: ˥ ˦ ˧ ˨ ˩, e.g., ɑ˥˩
• duration-characters which are considered: ˘ ˑ ː, e.g., ʌː
• normalize the text, e.g., numbers should be written out
• substituted space by either SIL0, SIL1 or SIL2 depending on the duration of the pause
  • use SIL0 for no pause
  • use SIL1 for a short pause, for example after a comma ...|v|i|ˈɛ|n|ʌ|,|SIL1|ˈɔ|s|t|ɹ|i|ʌ|...
  • use SIL2 for a longer pause, for example after a sentence: ...|ˈɝ|θ|.|SIL2
• Note: only phonemes occurring in the TextGrids (on the selected tier) are possible to synthesize

Synthesis

To prepare a text for synthesis, following things need to be considered:

• each line in the text file will be synthesized as a single file, therefore it is recommended to place each sentence onto a single line
• paragraphs can be separated by a blank line
• each symbol needs can be separated by an separator like |, e.g. s|ˌɪ|ɡ|ɝ|ˈɛ|t
  • this is useful if the model contains phonemes/symbols that consist of multiple characters, e.g., ˈɛ

Example valid sentence: "As the overlying plate lifts up, it also forms mountain ranges." => ˈæ|z|SIL0|ð|ʌ|SIL0|ˌoʊ|v|ɝ|l|ˈaɪ|ɪ|ŋ|SIL0|p|l|ˈeɪ|t|SIL0|l|ˈɪ|f|t|s|SIL0|ˈʌ|p|,|SIL1|ɪ|t|SIL0|ˈɔ|l|s|oʊ|SIL0|f|ˈɔ|ɹ|m|z|SIL0|m|ˈaʊ|n|t|ʌ|n|SIL0|ɹ|ˈeɪ|n|d͡ʒ|ʌ|z|.|SIL2

Example invalid sentence: "Digestion is a vital process which involves the breakdown of food into smaller and smaller components, until they can be absorbed and assimilated into the body." => daɪˈʤɛsʧʌn ɪz ʌ ˈvaɪtʌl ˈpɹɑˌsɛs wɪʧ ɪnˈvɑlvz ðʌ ˈbɹeɪkˌdaʊn ʌv fud ˈɪntu ˈsmɔlɝ ænd ˈsmɔlɝ kʌmˈpoʊnʌnts, ʌnˈtɪl ðeɪ kæn bi ʌbˈzɔɹbd ænd ʌˈsɪmʌˌleɪtɪd ˈɪntu ðʌ ˈbɑdi.

Pretrained Models

• LJS-IPA-101500: Model trained on LJ Speech dataset with IPA transcriptions for 101500 iterations (= 500 epochs) with separated learning of stress
  • Symbolset: ! " ' ( ) , - . : ; ? SIL0 SIL1 SIL2 [ ] aɪ aʊ b d d͡ʒ eɪ f h i j k l m n oʊ p s t t͡ʃ u v w z æ ð ŋ ɑ ɔ ɔɪ ɛ ɝ ɡ ɪ ɹ ʃ ʊ ʌ ʒ ˈaɪ ˈaʊ ˈeɪ ˈi ˈoʊ ˈu ˈæ ˈɑ ˈɔ ˈɔɪ ˈɛ ˈɝ ˈɪ ˈʊ ˈʌ ˌaɪ ˌaʊ ˌeɪ ˌi ˌoʊ ˌu ˌæ ˌɑ ˌɔ ˌɔɪ ˌɛ ˌɝ ˌɪ ˌʊ ˌʌ θ
• LJS-IPA-102000-durations: Model trained on LJ Speech dataset with IPA transcriptions for 102000 iterations (= 500 epochs) with separated learning of stress and phoneme durations
  • Symbolset: ! " ' ( ) , - — . : ; ? SIL0 SIL1 SIL2 SIL3 [ ] aɪ aɪː aɪˑ aɪ˘ aʊ aʊː aʊˑ aʊ˘ b bː bˑ b˘ d dː dˑ d˘ d͡ʒ d͡ʒː d͡ʒˑ d͡ʒ˘ eɪ eɪː eɪˑ eɪ˘ f fː fˑ f˘ h hː hˑ i iː iˑ i˘ j jː jˑ j˘ k kː kˑ k˘ l lː lˑ l˘ m mː mˑ m˘ n nː nˑ n˘ oʊ oʊː oʊˑ oʊ˘ p pː pˑ p˘ s sː sˑ s˘ t tː tˑ t˘ t͡ʃ t͡ʃː t͡ʃˑ t͡ʃ˘ u uː uˑ u˘ v vː vˑ v˘ w wː wˑ w˘ z zː zˑ z˘ æ æː æˑ æ˘ ð ðː ðˑ ð˘ ŋ ŋː ŋˑ ŋ˘ ɑ ɑː ɑˑ ɑ˘ ɔ ɔɪ ɔɪː ɔɪ˘ ɔː ɔˑ ɔ˘ ɛ ɛː ɛˑ ɛ˘ ɝ ɝː ɝˑ ɝ˘ ɡ ɡː ɡˑ ɡ˘ ɪ ɪː ɪˑ ɪ˘ ɹ ɹː ɹˑ ɹ˘ ʃ ʃː ʃˑ ʃ˘ ʊ ʊː ʊˑ ʊ˘ ʌ ʌː ʌˑ ʌ˘ ʒ ʒː ʒˑ ʒ˘ ˈaɪ ˈaɪː ˈaɪˑ ˈaɪ˘ ˈaʊ ˈaʊː ˈaʊˑ ˈaʊ˘ ˈeɪ ˈeɪː ˈeɪˑ ˈeɪ˘ ˈi ˈiː ˈiˑ ˈi˘ ˈoʊ ˈoʊː ˈoʊˑ ˈoʊ˘ ˈu ˈuː ˈuˑ ˈu˘ ˈæ ˈæː ˈæˑ ˈæ˘ ˈɑ ˈɑː ˈɑˑ ˈɑ˘ ˈɔ ˈɔɪ ˈɔɪː ˈɔɪˑ ˈɔɪ˘ ˈɔː ˈɔˑ ˈɔ˘ ˈɛ ˈɛː ˈɛˑ ˈɛ˘ ˈɝ ˈɝː ˈɝˑ ˈɝ˘ ˈɪ ˈɪː ˈɪˑ ˈɪ˘ ˈʊ ˈʊː ˈʊˑ ˈʊ˘ ˈʌ ˈʌː ˈʌˑ ˈʌ˘ ˌaɪ ˌaɪː ˌaɪˑ ˌaɪ˘ ˌaʊ ˌaʊː ˌaʊˑ ˌaʊ˘ ˌeɪ ˌeɪː ˌeɪˑ ˌeɪ˘ ˌi ˌiː ˌiˑ ˌi˘ ˌoʊ ˌoʊː ˌoʊˑ ˌoʊ˘ ˌu ˌuː ˌuˑ ˌu˘ ˌæ ˌæː ˌæˑ ˌæ˘ ˌɑ ˌɑː ˌɑˑ ˌɑ˘ ˌɔ ˌɔɪ ˌɔɪː ˌɔɪˑ ˌɔɪ˘ ˌɔː ˌɔˑ ˌɔ˘ ˌɛ ˌɛː ˌɛˑ ˌɛ˘ ˌɝ ˌɝː ˌɝˑ ˌɝ˘ ˌɪ ˌɪː ˌɪˑ ˌɪ˘ ˌʊ ˌʊː ˌʊˑ ˌʊ˘ ˌʌ ˌʌː ˌʌˑ ˌʌ˘ θ θː θˑ θ˘

Audio Example

"The North Wind and the Sun were disputing which was the stronger, when a traveler came along wrapped in a warm cloak." Listen here (headphones recommended)

Example Synthesis

To reproduce the audio example from above, you can use the following commands:

# Create example directory
mkdir ~/example

# Download pre-trained Tacotron model checkpoint
wget https://tuc.cloud/index.php/s/xxFCDMgEk8dZKbp/download/LJS-IPA-101500.pt -O ~/example/checkpoint-tacotron.pt

# Download pre-trained Waveglow model checkpoint
wget https://tuc.cloud/index.php/s/yBRaWz5oHrFwigf/download/LJS-v3-580000.pt -O ~/example/checkpoint-waveglow.pt

# Create text containing phonetic transcription of: "The North Wind and the Sun were disputing which was the stronger, when a traveler came along wrapped in a warm cloak."
cat > ~/example/text.txt << EOF
ð|ʌ|SIL0|n|ˈɔ|ɹ|θ|SIL0|w|ˈɪ|n|d|SIL0|ˈæ|n|d|SIL0|ð|ʌ|SIL0|s|ˈʌ|n|SIL0|w|ɝ|SIL0|d|ɪ|s|p|j|ˈu|t|ɪ|ŋ|SIL0|h|w|ˈɪ|t͡ʃ|SIL0|w|ˈɑ|z|SIL0|ð|ʌ|SIL0|s|t|ɹ|ˈɔ|ŋ|ɝ|,|SIL1|h|w|ˈɛ|n|SIL0|ʌ|SIL0|t|ɹ|ˈæ|v|ʌ|l|ɝ|SIL0|k|ˈeɪ|m|SIL0|ʌ|l|ˈɔ|ŋ|SIL0|ɹ|ˈæ|p|t|SIL0|ɪ|n|SIL0|ʌ|SIL0|w|ˈɔ|ɹ|m|SIL0|k|l|ˈoʊ|k|.|SIL2
EOF

# Synthesize text to mel-spectrogram
tacotron-cli synthesize \
  ~/example/checkpoint-tacotron.pt \
  ~/example/text.txt \
  --sep "|"

# Install waveglow-cli for synthesis of mel-spectrograms
pip install waveglow-cli --user

# Synthesize mel-spectrogram to wav
waveglow-cli synthesize \
  ~/example/checkpoint-waveglow.pt \
  ~/example/text -o

# Resulting wav is written to: ~/example/text/1-1.npy.wav

Roadmap

• Outsource method to convert audio files to mel-spectrograms before training
• Better logging
• Provide more pre-trained models
• Add audio examples
• Adding tests

License

MIT License

Acknowledgments

Model code adapted from Nvidia.

Papers:

• Tacotron: Towards End-to-End Speech Synthesis
• Natural TTS Synthesis by Conditioning Wavenet on MEL Spectrogram Predictions

Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project-ID 416228727 – CRC 1410

Citation

If you want to cite this repo, you can use this BibTeX-entry generated by GitHub (see About => Cite this repository).