| Property | Value |
|---|---|
| Parameter Count | 88M |
| Architecture | BERT-base |
| Max Sequence Length | 512 |
| Training Data | 330M phoneme-level sentences |
| Paper | INTERSPEECH 2023 |
What is xphonebert-base?
XPhoneBERT is a groundbreaking pre-trained multilingual model specifically designed for phoneme representations in text-to-speech (TTS) applications. Built on the BERT-base architecture and trained using RoBERTa's pre-training approach, it processes phoneme-level sentences across approximately 100 languages and locales.
Implementation Details
The model leverages the transformers library and requires the text2phonemesequence package for converting text into phoneme-level sequences. It employs specialized word segmentation and text normalization techniques, utilizing tools like spaCy and VnCoreNLP for different languages.
- Incorporates CharsiuG2P and segments toolkits for text-to-phoneme conversion
- Supports ISO 639-3 language codes for multiple languages
- Implements BERT-base architecture with 88M parameters
- Maximum sequence length of 512 tokens
Core Capabilities
- Multilingual phoneme representation generation
- Enhanced naturalness and prosody in TTS systems
- Effective performance with limited training data
- Support for nearly 100 languages and locales
Frequently Asked Questions
Q: What makes this model unique?
XPhoneBERT is the first pre-trained multilingual model specifically designed for phoneme representations in TTS. Its ability to work across nearly 100 languages and significantly improve TTS quality sets it apart from other models.
Q: What are the recommended use cases?
The model is ideal for text-to-speech applications, especially in scenarios requiring high-quality multilingual speech synthesis, prosody enhancement, or when working with limited training data.