Unverified85% confidenceFactTime unknown
VALL-E、SoundStorm等主流大模型动辄数十亿参数
1
Sources
85%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhisper、SoundStorm、AudioLM等语音基础模型推动了端到端范式走向成熟57% similarUnverifiedNVIDIA has launched the AUDEX-310 audio unified model, with a total parameter count of 30 billion but only 3 billion activated parameters, adopting a sparse activation architecture and supporting speech recognition, translation, and sentiment analysis.57% similarUnverified自回归模型（如VALL-E）逐token预测语音离散表征，能建模长程依赖和丰富韵律，但存在累积误差；非自回归模型（如FastSpeech）并行生成速度快但表现力受限56% similarUnverifiedThe effective combinatorial space of multi-layer audio codebooks can reach billions, far exceeding the text vocabulary which is typically on the order of tens of thousands to hundreds of thousands, increasing the difficulty of safety alignment.55% similarUnverifiedSpeech synthesis models such as VALL-E and VoiceCraft adopt the technical approach of representing speech as token sequences and using an autoregressive generation framework.54% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/19915API
curl https://kongchang.com/api/v1/knowledge/claims/19915MCP
get_claim(id=19915)