Commit History

2 days ago

  1. Fill a take: 800 chars is ~54s, just under the frame cap, instead of stopping at 600 Running

    VoiceClimax
    78675ec
  2. Split a take into start (GPU, charged to the caller's IP) and stream (direct)

    VoiceClimax
    3c0fa66
  3. History turns may carry audio instead of codes, for backends that cannot export frames

    VoiceClimax
    23534c2
  4. Accept vq_codes references so a context token skips the codec

    VoiceClimax
    d96f284
  5. Add /speech: chunked take with a context token, and emit the sampled codes

    VoiceClimax
    a0646ab
  6. Keep local notes out of the published tree

    VoiceClimax
    4875b2e
  7. Point the preset at the clip we ship; skip missing presets instead of failing startup

    VoiceClimax
    abc57c7
  8. Streaming speech service on ZeroGPU

    VoiceClimax
    b1c28a5