dsh-hold-to-talk: Local Hold-to-Talk Voice Input for the DeepSeek Harness Web Input Field
dsh-hold-to-talk is a local, offline voice input plugin for the DeepSeek Harness Web composer. It replicates WeChat’s “hold-to-talk” interaction by invoking the SenseVoice model in the browser, requiring no API key and keeping all audio processing local. The plugin supports long-press recording, real-time preview, swipe-up cancellation, and automatic completion; when recording ends, the transcribed text is automatically inserted into the draft. The initial use requires downloading a model of approximately 230 MB. It requires a Chromium-based browser and a Node.js runtime, making it suitable for privacy- and data-security-focused local deployment scenarios.
Read Moredsh-voice-funasr: DSH Web UI local offline speech input plugin
dsh-voice-funasr is a local offline speech input plugin for DSH Web UI, designed to provide a speech transcription solution that does not rely on cloud speech services. Its core mechanism is "local-first," prioritizing local audio transcription using the FunASR int8 ONNX engine (including Paraformer, FSMN-VAD, and ct-punc models). If the local engine is unavailable, it automatically falls back to the browser's Web Speech API. The plugin supports a "hold to record, release to send" operation and optionally invokes a user-configured LLM endpoint for two-step polishing to correct stutters or filler words; this is the only component that may send data over the network, while all other data is processed locally in memory. Installation requires installing funasr-onnx and other dependencies via pip, and downloading...
Read More