Unverified50% confidenceReleaseExact time
Block released DESA, a speech model that supports local operation, can listen and speak in real time and can interject and respond, reportedly trained from scratch using only human speech without relying on text training or pretrained models.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/25/2026
First Seen
Valid until: 1/23/2027
Sources
Related Entities
Related Claims
UnverifiedThe speech model DESA released by a Twitter-affiliated team uses no text pretraining at all, learning to speak by listening to real human conversations, features full-duplex interruption, breathing and laughter, and can run locally on a single Apple computer.74% similarUnverified零样本迁移学习通过大规模预训练学会从极短音频中提取说话人嵌入向量，无需针对目标说话人专项训练即可复刻音色语调节奏72% similarUnverifiedVALL-E, released by Microsoft in 2023, models speech generation as a language model problem, requiring only 3 seconds of reference audio to achieve zero-shot cloning, and uses the EnCodec neural audio codec.70% similarUnverifiedVoiceCap promises that recording data will not be used to train models.70% similarVerifiedModels such as Whisper and Wav2Vec 2.0 learn general acoustic representations on large amounts of unlabeled audio through self-supervised pre-training, reducing reliance on manually annotated data.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1095774API
curl https://kongchang.com/api/v1/knowledge/claims/1095774MCP
get_claim(id=1095774)