TunArTTS: Tunisian Arabic Text-To-Speech Corpus

Imen Laouirine, Rami Kammoun, Fethi BougaresCOLING 2024

TL;DR

This paper introduces TunArTTS, the first publicly available text-to-speech (TTS) corpus for the Tunisian Arabic dialect, addressing the significant gap in low-resource TTS research. By leveraging transfer learning from English, they achieved a promising Mean Opinion Score (MOS) of 3.88, paving the way for more natural and accessible Tunisian TTS systems.

What Problem Does This Paper Address?

Text-to-Speech (TTS) technology, the ability to convert written text into spoken audio, has seen remarkable advancements, particularly for high-resource languages like English. However, many languages, especially Arabic dialects, remain underserved due to a lack of sufficient annotated speech data. The Tunisian dialect, in particular, has historically lacked any prior TTS research or publicly available datasets. This absence hinders the development of natural-sounding and intelligible TTS systems for millions of Tunisian speakers, limiting their access to voice-enabled technologies and digital content. This paper aims to bridge this gap by creating a foundational dataset and demonstrating initial TTS capabilities for the Tunisian dialect.

Methods

The authors explored two primary approaches for building their Tunisian TTS system:

  • Training from Scratch: This involved training acoustic models (specifically Tacotron2) and vocoders (Parallel WaveGAN) entirely on the newly created TunArTTS dataset.
  • Transfer Learning: This approach leveraged pre-trained acoustic models from other languages (English, Arabic, and French) and fine-tuned them on the TunArTTS dataset. This is a common strategy for low-resource languages, as pre-trained models have already learned general speech synthesis principles.

They experimented with different pre-trained models, including Tacotron2 and FastSpeech2, and evaluated various hyperparameter settings. The vocoder stage utilized Parallel WaveGAN, chosen for its clear audio output. For acoustic model training, they explored both character-based and phonetic-based representations, with the latter proving more effective.

Datasets

  • TunArTTS (New Benchmark): The core contribution is the creation of the TunArTTS corpus. It was sourced from an online English and Tunisian Arabic dictionary, yielding over 3 hours of mono-speaker (male) audio sampled at 44.1 kHz. The dataset was manually diacritized and cleaned.
  • LJSpeech: A widely used English TTS dataset (24 hours of a single female speaker) used for pre-training models in the transfer learning experiments.
  • SIWIS: A French TTS dataset (10 hours) used for pre-training.
  • Arabic Speech Corpus (ASC) and CIArTTS: Arabic datasets used for pre-training, with ASC showing better results than CIArTTS.

Key Insights and Ideas

  • Transfer Learning is Crucial for Low-Resource Dialects: The paper strongly demonstrates that transfer learning, particularly from English, significantly outperforms training from scratch for the Tunisian dialect. This highlights the effectiveness of leveraging knowledge from high-resource languages to bootstrap TTS development in under-resourced ones.
  • English as a Strong Source for Transfer: The best results were achieved by fine-tuning a model pre-trained on the English LJSpeech dataset. This suggests that the general acoustic and prosodic patterns learned from English can be effectively adapted to Tunisian Arabic, even with its distinct linguistic features.
  • Diacritization is Essential for Tunisian Arabic TTS: The paper emphasizes the necessity of diacritics (vowel markings) for accurate pronunciation and disambiguation in written Tunisian Arabic, a feature often omitted in everyday use but critical for TTS systems.
  • Dataset Quality Matters: Despite the TunArTTS dataset being derived from a dictionary and not specifically designed for TTS, careful cleaning, diacritization, and selection of audio segments were vital for achieving intelligible speech.

Key Results

  • Best MOS Score: The best performing model, achieved through transfer learning from an English pre-trained Tacotron2 model, obtained a Mean Opinion Score (MOS) of 3.88. This is a significant achievement for a low-resource dialect.
  • Comparison to Training from Scratch: The transfer learning model significantly outperformed the model trained from scratch, which achieved a MOS of 2.86.
  • Objective Evaluation: While subjective MOS scores showed a clear advantage for transfer learning, objective metrics like MCD, WER, and CER also indicated better performance for the English transfer learning model compared to training from scratch, though the differences were less pronounced.
  • Dataset Size: The TunArTTS corpus contains approximately 3 hours and 32 seconds of audio, with an average clip duration of 7.24 seconds.

Limitations and Future Work

  • Mono-Speaker Dataset: The current TunArTTS dataset consists of only one male speaker. This limits the diversity of voices and the ability to generate speech in different styles or genders.
  • Potential for Annotation Improvement: The authors acknowledge that some French loanwords in Tunisian Arabic might require further refinement in annotation to preserve their original tone.
  • Future Directions:
    • Multi-Speaker Corpus: The team plans to expand the dataset by extracting and annotating audio from other speakers to create a multi-speaker Tunisian dialect TTS system.
  • Manual Text Revision: They intend to manually revise the text to be CODA-compliant (Conventionalized Orthography for Dialectal Arabic) to improve phonetic representation.
  • Exploring Other Dialects: The methodology could potentially be extended to other Arabic dialects.

Summary updated

BibTeX
@inproceedings{laouirine2024tunartts,
  author = {Imen Laouirine and Rami Kammoun and Fethi Bougares},
  title = {{TunArTTS: Tunisian Arabic Text-To-Speech Corpus}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}
Figure