0x16e8bea/protoly-cerebras-voice

Repair tool-message pairing per request: orphaned/duplicate outputs re-homed, missing outputs synthesized

Cut speculative LLM requests: raise VAD min_silence to 500ms, drop live transcription

Fix Qwen3-TTS boot: unify HF cache and pre-fetch full snapshot

Switch TTS from Kokoro to Qwen3-TTS CustomVoice

Retry Cerebras 429s with backoff and speak a rate-limit fallback

Restore responses adapter with tool forwarding

Document Cerebras tool-call backend

Use Cerebras chat completions for tool calls

Document Gemma Kokoro realtime setup

Use Gemma fallback model in start script

Use Gemma Cerebras model by default

Install Kokoro TTS dependencies

Switch proof run to Kokoro TTS

Use startup script

Add startup script

Copy adapter into Docker image

Run Cerebras responses adapter

Add Cerebras responses adapter

Remove unsupported qwen backend flag

Use responses-api backend flags

Use virtualenv for speech-to-speech install

Add Space metadata and notes

Add speech-to-speech Dockerfile

initial commit