Recent Advances in Discrete Speech Tokens: A Review.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Speech Technology

Background:

• Large language models (LLMs) have spurred advancements in speech generation.
• Discrete speech tokens offer efficient and compatible speech representation for LLMs.
• Current research divides discrete speech tokens into acoustic and semantic categories.

Purpose of the Study:

• To systematically synthesize the taxonomy and innovations in discrete speech tokenization.
• To critically examine the strengths and limitations of acoustic and semantic token paradigms.
• To provide experimental comparisons and identify future research directions.

Main Methods:

• Systematic literature review and synthesis of discrete speech tokenization research.
• Categorization of existing discrete speech token approaches.
• Experimental comparison of different token types.

Main Results:

• Discrete speech tokens are crucial for integrating speech into LLM architectures.
• Acoustic and semantic tokens represent distinct approaches with unique advantages and disadvantages.
• Experimental comparisons highlight performance variations across token types.

Conclusions:

• Discrete speech tokens are foundational for modern speech generation within LLMs.
• Further research is needed to address challenges and optimize token design and application.
• This survey provides insights for future advancements in speech tokenization technology.