A Language Modeling Approach to Diacritic-Free Hebrew TTS (Interspeech 2024)

Inference code and model weights for the paper "A Language Modeling Approach to Diacritic-Free Hebrew TTS" (Interspeech 2024).

---

Abstract: We tackle the task of text-to-speech (TTS) in Hebrew. Traditional Hebrew contains Diacritics (`Niqqud'), which dictate the way individuals should pronounce given words, however, modern Hebrew rarely uses them. The lack of diacritics in modern Hebrew results in readers expected to conclude the correct pronunciation and understand which phonemes to use based on the context. This imposes a fundamental challenge on TTS systems to accurately map between text-to-speech. In this study, we propose to adopt a language modeling Diacritics-Free TTS approach, for the task of Hebrew TTS. The language model (LM) operates on discrete speech representations and is conditioned on a word-piece tokenizer. We optimize the proposed method using in-the-wild weakly supervised recordings and compare it to several diacritic based Hebrew TTS systems. Results suggest the proposed method is superior to the evaluated baselines considering both content preservation and naturalness of the generated speech.

Try it out!

You can try our model in the google colab demo.

Installation

git clone https://github.com/slp-rl/HebTTS.git

We publish our checkpoint in google drive. AR model trained for 1.2M steps and NAR model for 200K steps on HebDB.

gdown 11NoOJzMLRX9q1C_Q4sX0w2b9miiDjGrv

Install Dependencies

pip install torch torchaudio
pip install torchmetrics
pip install omegaconf
pip install git+https://github.com/lhotse-speech/lhotse
pip install librosa
pip install encodec
pip install phonemizer
pip install audiocraft  # optional

Inference

You can play with the model with different speakers and text prompts.

run infer.py:

python infer.py  --checkpoint checkpoint.pt --output-dir ./out --text "היי מה קורה"

you can specify additional arguments --speaker and --top-k.

Multi Band Diffusion

Tip

We allow using the new Multi Band Diffusion (MBD) vocoder for generating a better quallity audio. Install audiocraft and set --mbd True flag.

Text

you can concatenate text prompts using | or specify a path of a text file spereated by \n if writing Hebrew in terminal is inconvenient.

תגידו גנבו לכם פעם את האוטו ופשוט ידעתם שאין טעם להגיש תלונה במשטרה
היי מה קורה
בראשית היתה חללית מסוג נחתת

and run

python infer.py  --checkpoint checkpoint.pt --output-dir ./out --text example.txt

Speakers

you can use the speaker defined in speakers.yaml, or add additional speakers. specify wav files and transcription in same format.

--speaker shaul

Citation

@article{roth2024language,
  title={A Language Modeling Approach to Diacritic-Free Hebrew TTS},
  author={Roth, Amit and Turetzky, Arnon and Adi, Yossi},
  journal={arXiv preprint arXiv:2407.12206},
  year={2024}
}

Acknowledgments

• Model code inside valle is based on the implementation of Feiteng Li.