Pierce Freeman and Richard Diehl-Martinez discuss demonstrations of Block’s DESA speech model and the difference between a speech-to-speechA speech-to-speech model takes spoken audio as input and generates spoken audio as output. Its design can model acoustic and conversational information beyond a plain transcript. approach and a pipeline that passes through text. They emphasize that the model and technical details were not available for their own independent testing.
Pierce Freeman and Richard Diehl-Martinez listen to a sample whose fluent delivery includes invented personal detailsA hallucination is an AI output that presents false, unsupported, or invented information as though it were reliable. and a story that becomes inconsistent. They discuss how realistic voice and conversational timing may make such errors less noticeable, as well as concerns about the selected example.
The hosts compare another demonstration with a conversational model and speculate about interruption, training data and audio generationAudio generation uses AI to create speech, music, sound effects or other audio from instructions and reference inputs.. Their proposed explanations are hypotheses rather than verified descriptions of DESA’s implementation.
Watch on YouTube
