Trelis/duplex-pre-training-yodas3-mimi
Viewer • Updated • 3.63k • 1.6k • 32
Duplex speech-to-speech: tokenized English speech datasets with separated speaker tracks, and the duplex-0 model.
Note Pre-training: 408 hours from YODAS3.
Note Pre-training: 340 hours from podcasts.
Note Conversational: 156 hours of two-speaker podcast conversation with frequent turn-taking.
Note duplex-0: a 201M-parameter duplex speech-to-speech research model.