Unverified50% confidenceBenchmarkExact time
Speech generated by end-to-end TTS models based on deep learning typically achieves a naturalness MOS score of 4.0–4.5 (out of 5), approaching the level of human recordings.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/15/2026
First Seen
Valid until: 11/13/2026
Sources
PPT Master：AI一键生成原生可编辑PowerPoint演示文稿
Related Claims
UnverifiedThe new end-to-end voice model is trained directly at the audio level, without first converting speech to text, and can directly capture acoustic features and emotions.70% similarUnverified声音克隆通过说话人编码器将声音映射为高维嵌入向量，说话人编码器训练通常采用对比学习或三元组损失70% similarUnverified语音数据通常被编码为离散的音频token，每秒语音大约对应25-50个token69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/752101API
curl https://kongchang.com/api/v1/knowledge/claims/752101MCP
get_claim(id=752101)