dsh-speech

DSH's official host has no speech seam at all: the LLM extension point only understands chat modality, and the attachment channel accepts images exclusively. dsh-speech mounts an independent speech service surface on the standard webServer route registration extension: /s/api for batch ASR and TTS, /s/ws for a real-time transcription WebSocket session channel. Phone apps or any LAN client can call directly with a single shared Bearer token—no DSH source changes, no recompilation, no re-download.

Four adapter families power the core capability layer: Alibaba DashScope native protocol (batch ASR/TTS plus sentence-level real-time transcription), the OpenAI-compatible API (OpenAI, Groq, SiliconFlow, or any custom gateway—switching vendors is a config-only change), direct local service relay (self-hosted SenseVoice and CosyVoice2 with batch and streaming synthesis plus voiceprint speaker diarization), and streaming WebSocket upstreams (Deepgram, FunASR, sherpa-onnx, iFlytek IAT/RTASR/TTS). The three selectors for batch ASR, synthesis, and session transcription are fully independent, allowing arbitrary mixes such as DashScope recognition plus local synthesis plus Deepgram real-time. Real-time transcription supports VAD sentence segmentation, pseudo-streaming word-level previews, local-embedding speaker diarization, and iFlytek long-form multi-speaker real-time transcription (roleType=2) with multi-dialect online synthesis. A built-in model-fetch button in the editor refreshes the catalog with one click when providers rename models—no code changes needed.

Ideal for teams that need speech input and audio output for mobile or LAN desktop clients; developers who must orchestrate multiple cloud and local speech models in a single pipeline; and anyone who wants a complete ASR/TTS/real-time transcription stack on their DSH host with a single plugin install—no forking, no recompilation.