CoolFace
Apppublic

0x16e8bea/protoly-cerebras-voice

sourceHugging Faceupdated 3mo agoView on Hugging Face
0likes
App README

Protoly Cerebras Voice

Docker Space for testing a realtime voice backend with ASR/TTS hosted on Hugging Face GPU hardware and text generation routed to Cerebras.

Runtime shape:

  • —Realtime WebSocket: /v1/realtime
  • —ASR: Parakeet TDT
  • —TTS: Qwen3-TTS (Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice, speaker Aiden)
  • —LLM: Cerebras Chat Completions via the local Responses API adapter
  • —Tool calls: forwarded through the adapter and returned as OpenAI Realtime-compatible tool-call events for app-side delegation

Required secret:

  • —CEREBRAS_API_KEY

Optional variables:

  • —CEREBRAS_MODEL defaults to gemma-4-31b
  • —CEREBRAS_BASE_URL defaults to https://api.cerebras.ai/v1
  • —CEREBRAS_RETRY_BUDGET_S defaults to 14 (in-adapter 429 retry window)
  • —CEREBRAS_RATE_LIMIT_FALLBACK spoken text when the rate limit persists
  • —QWEN3_TTS_MODEL defaults to Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
  • —QWEN3_TTS_SPEAKER defaults to Aiden
  • —VAD_MIN_SILENCE_MS defaults to 500 (lower = snappier turn ends but more speculative LLM requests against the Cerebras quota)