Generative Spoken Language Models Achieve Quality Speech at Lower Bitrates

A new research paper published on arXiv explores the effectiveness of Generative Spoken Language Modeling (GSLM) for speech synthesis and continuation. The study investigates how varying segmentation widths and cluster sizes in k-means clustering affect speech quality at different bitrates. Researchers found that intelligible and natural speech can be achieved at lower bitrates than previously thought, with speech continuation quality remaining stable. The paper suggests that current GSLM settings might be unnecessarily complex and highlights the need for improved automatic evaluation methods, as LLM-based metrics still show low correlation with human subjective scores. AI

IMPACT This research could lead to more efficient and less computationally intensive models for speech synthesis and continuation.

RANK_REASON Academic paper detailing a new methodology and findings in speech synthesis. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Generative Spoken Language Modeling
• k-means clustering

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →