---

name: components-and-backends description: "Choose, configure, and troubleshoot speech-to-speech VAD, STT, LLM, and TTS backends, optional extras, language behavior, prompts, tools, and benchmark safety." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

Components and backends

Use this sub-skill when the task is to choose or debug a VAD, STT, LLM, or TTS component for speech-to-speech, including optional extras, model/device choices, Qwen3-TTS modes, direct audio-input LLMs, language behavior, prompts, tool calls, and benchmark planning.

Route by question

• Need a supported backend list or install extra: read Backend catalog.
• Need a concrete model, language, direct-audio, Qwen3, or Apple Silicon recipe: read Model and language recipes.
• Need LLM prompt or tool-calling behavior: read Prompting and tools.
• Need latency/RTF comparison or a benchmark plan: read Backend benchmarking before running any model-heavy measurement.
• Need optional dependency, CUDA wheel, model cache, language, or voice troubleshooting: read Troubleshooting.
• Need command shape, serve/local/talk, host/port, VAD flags, or session capacity: route to cli-and-server.
• Need Realtime event lifecycle, WebSocket/WebRTC, or client protocol: route to realtime-api.
• Need the browser demo tools/search/camera/UI: route to browser-demo.

Safe operating sequence

1. Select one backend per stage with --stt, --llm_backend, and --tts. Defaults are parakeet-tdt, responses-api, and qwen3.
2. Install only the optional extra required by the selected backend. Do not install every extra just to try one component.
3. Verify parser/config behavior before model inference. Run the bundled safe registry inspector: scripts/inspect_backend_registry.py.
4. For local model inference, decide the hardware/backend first: CPU/CUDA, Apple Silicon/MPS/MLX, remote OpenAI-compatible endpoint, or direct audio-input Chat Completions model.
5. For Qwen3-TTS on Linux, resolve the qwentts-cpp-python wheel/runtime match before diagnosing model code. For Apple Silicon, use the MLX quantization mapping described in the recipes.
6. Treat full STT/TTS benchmarks and live speech loops as optional, expensive validation. They may download models, require CUDA/MPS/audio devices, or use credentials.

High-signal rules

• --stt none bypasses transcription and forwards the completed VAD audio turn to an LLM that supports audio input. In this package that means --llm_backend chat-completions plus an explicitly audio-capable --model_name. Do not pair --stt none with the default responses-api profile.
• responses-api and chat-completions share the --responses_api_* connection flags. They call different upstream endpoints.
• Local transformers and mlx-lm tool calling uses prompt-rendered function signatures and parsed <code>...</code> calls; remote OpenAI-compatible backends pass tools structurally.
• --mac-optimal-settings supplies macOS defaults only. Any explicit backend, model, global device, or component-device flag wins over the preset.
• Backend registry import/config checks do not prove that a large model can run. Record actual model inference separately when a task requires it.