Smooth playback and guard visitor role

Stream guarded stable audio prefixes

Ground Lyra and drain completed replies

Measure RoleForge latency and streaming equivalence

Allow longer bounded LFM story replies

Extend bounded LFM story replies

Reject repeated generic LFM openings

Allow complete long LFM replies

Keep LFM text and audio aligned

Reduce coherent LFM turn latency

Honor the official LFM interleaved contract

Make live dialogue quality-first and server-gated

Restore sample-exact browser playback

Ground Lyra and regenerate early repeats

Stabilize PCM playback and barge-in

Attach PCM routes to Gradio server

Replace Gradio audio with WebSocket PCM transport

Smooth live audio streaming

Start live session from microphone recording

Repair live microphone session routing

Add interruptible live LFM conversation

Share LFM context across ZeroGPU workers

Retain bounded LFM conversation context

Migrate RoleForge feasibility lab to LFM2.5-Audio

Apply TURN relay to both WebRTC peers

Buffer ICE without changing WebRTC component identity

Revert "Buffer early WebRTC ICE candidates"

Buffer early WebRTC ICE candidates

Warm PersonaPlex under live inference mode

Relay Space WebRTC through short-lived TURN

Spike PersonaPlex WebRTC streaming on ZeroGPU

Repair microphone capture frontend

Use ZeroGPU supported torch version syntax

Repair PersonaPlex for Blackwell ZeroGPU

Pin Gradio-compatible web stack

Pin Pydantic for Gradio schema compatibility

Align Gradio with PersonaPlex dependencies

Use platform-managed ZeroGPU runtime

Integrate generated Hugging Face Space scaffold

initial commit

Use scoped OAuth for gated model access

Build RoleForge voice NPC feasibility lab

Add support for audio input upload in demo interface

Enhance audio processing in generate_response by adding prepend and append silence to improve model response timing

Add warmup function to initialize CUDA graphs and improve performance in get_models

Refine sine wave token usage in generate_response function to simulate user silence with official PersonaPlex constants

Reset streaming state within context in generate_response function

Enhance response generation by adding silence continuation and simulating user input with sine wave tokens

Reorganize imports for moshi models and streamline text tokenizer loading in get_models function

Implement lazy loading for models in get_models function to optimize GPU resource usage