Unverified70% confidenceFactExact time
AudioPaLM, released by Google in 2023, was the first to inject speech tokens directly into a language model's vocabulary, enabling joint modeling of text and audio in the same representation space.
2
Sources
70%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
Unverified端到端语音架构的早期探索可追溯到2014年前后Google提出的Sequence-to-Sequence模型，以及2016年百度提出的Deep Speech 278% similarUnverifiedGoogle's previous speech recognition capabilities were primarily provided in the form of the Cloud Speech-to-Text API, oriented toward backend integration for enterprises and developers.74% similarUnverified工具增强语言模型（Tool-Augmented Language Model，TALM）最早的系统性论述可追溯至 2022 年 Google 发表的同名论文70% similarUnverifiedGoogle possesses the world's largest accumulation of multilingual speech data from YouTube, search voice input, and the Android system, constituting a moat at the training data level.70% similarUnverifiedGoogle launched the BERT pre-trained language model in 2018 and the PaLM model with 540 billion parameters in 2022.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/520896API
curl https://kongchang.com/api/v1/knowledge/claims/520896MCP
get_claim(id=520896)