The native view follows the inspected NVIDIA generator, codebook timing, codec and reference client. The connected views show alternative computational paths toward speaking while Inkling continues work.
What we can reuse
PersonaPlex’s concurrent audio input/output, role and voice prompts, temporal model, depth decoder and Mimi. Existing Inkling states, grounded speech text and source audio can be reused to prepare training examples.
Earlier connector proposals
The Single connector, Four connectors and Semantic outputs tabs preserve earlier designs. The selected Sequence memory tab documents the implemented Stage 1 memory writer and readers; it supersedes those proposals for the current training run.
Single connector
One connection before temporal block 16, width 512, with 8.40M parameters. Begin with PersonaPlex frozen; consider temporal LoRA only after observing a need. Dimensions and placement are hypotheses.
Four connectors
Four independent connections before blocks 4, 12, 20 and 28, each width 2048: 134.25M connector parameters. After initial alignment, jointly fine-tune the temporal model, final normalization and text head. The proposed audio return captures rendered samples, resamples to 16 kHz, and reuses Inkling’s native dMel and audio embeddings. Its assistant-audio wrapper is unvalidated. Tool-result text encoding remains open.
Semantic outputs
PersonaPlex starts independently while Inkling continuously processes native audio. Selected states arrive layer by layer; no full-forward or all-selected-layer wait is required. The proposal keeps all 32 temporal blocks and computes one shared attention update after PersonaPlex’s final normalization. Separate manual coefficients route that update to the native text head, the first audio-code input, or both. Later audio-code inputs retain the original temporal state. No agent audio is returned to Inkling.
Inference-only comparison
The semantic-output view specifies a closed reference, text-only, first-code-only and combined condition with fixed weights. Its first codebook is motivated by Mimi’s semantic-token experiment; content, voice and timing are not assumed to be perfectly separated. Garbled and muffled output remains relevant evidence.
Current sequence memory · Stage 1
Implemented training architecture: three Inkling rows (28, 34, final 42) per token, one shared 1024-wide writer and four independent 32-head readers after temporal blocks 7, 15, 23 and 31. A rolling 40,000-token memory combines hidden states, whole-word text features, token kind, recency and elapsed time. There is no end connector or memory time rotation. The 184,691,712 writer/reader parameters train with weighted text and all 16 voice-codebook losses; Inkling, PersonaPlex and Mimi stay frozen.
Parallel conversation
PersonaPlex can be partway through an explanation while Inkling continues work. New information may affect future generated frames. Thinking tokens, audio tokens and audible words have different rates; neither model is assumed to run ahead.
The long-term observation
New information changes an ongoing explanation appropriately while the agent keeps working. Early untrained probes may be rough: name evidence, response changes and timing should be reported separately from fluency. The semantic-output path is not implemented or measured.
Model dimensions were checked in the specified source version; checkpoint tensors were not loaded. Native benchmark figures do not establish performance of this integration. The released server has no input for tool results or Inkling states during a call.