Fix WebSocketDisconnect tracebacks and handle closed receives gracefully

Fix VAD queue corruption, ASR duplicate calls, and LLM TTFT bottlenecks

Fix NameError: AVAILABLE_VOICES not defined by reordering variables

Optimize TTFT, add connection pooling, implement streaming ASR with stabilization and waterfall metrics

Refine Stage 2: Add MemoryService, JSON logging middleware, and fuzzed stress tests

Deploy Stage 2 decoupled Event-Driven Conversational Runtime architecture

Optimize transcription quality: set Whisper temperature=0.0 and increase VAD voice sensitivity (minimum threshold 0.008)

Fix premature done-cutoff of audio queue and add Voice Interruption UI toggle checkbox

Fix premature cut-off and echo feedback: ignore mic bytes while responding, increase VAD silence threshold to 0.8s, increase client interruption threshold to 0.06

Optimize speech-to-speech interaction: responsive VAD, concurrent tasks, async clients, and lookahead jitter buffer

Optimize VAD silence duration to 1.0s and expand pre-speech buffer size to 4

Support streaming on /tts GET/POST endpoints and fix VAD transcription word filters

Optimize build speed: cache CPU-only PyTorch and remove custom USER instruction

Security Hardening: CORS limits, WebSocket token auth & rate limits, audio size caps, prompt sanitization, error message masking, and non-root Docker user

Update S2S server and client interface

Optimize Whisper transcription with accent-mapping prompts and add UI language/accent selector

Expose standalone TTS REST endpoints (GET/POST /tts) and document them

feat: STT accuracy, persona fine-tuning, voice REST API, updated UI and docs

Fix Whisper hallucinations: dynamic noise-floor VAD, 2-frame debounce, and expanded filter

Set default PORT to 7860 for Hugging Face Spaces

Deploy Speech-to-Speech real-time app to HF Spaces

initial commit