Jaehan

Reliable speech recognition depends on matching encoders, decoders, speaker models, and customization tools to the audio and deployment constraints of each application.

• What must a speech system handle?
  0:32 ↗
• Choose the decoder for the workload
  2:42 ↗
• Shorten the sequence before decoding
  4:26 ↗
• Parakeet and Canary divide the work
  5:27 ↗
• Connect speaker timestamps to transcript tokens
  6:57 ↗
• Make recognized speech usable
  8:26 ↗
• Recognize lyrics over music
  9:50 ↗
• Build coverage into the training data
  10:53 ↗
• Keep training fed and validation broad
  12:01 ↗
• Serve streaming and offline workloads
  13:22 ↗
• Adapt the layer responsible for the error
  14:27 ↗
• Find a model, then deploy or fine-tune it
  15:32 ↗