dsh-cyberdog-speech-sherpa: DeepSeek Harness Local Offline Speech Plugin
dsh-cyberdog-speech-sherpa is a local offline voice plugin for DeepSeek Harness (DSH) based on Sherpa-ONNX, designed to address voice input and output requirements in offline or intranet environments. Running entirely locally without relying on online services and compatible with CPU, the plugin primarily offers three core capabilities: first, Chinese speech-to-text (STT) using the Zipformer model; second, Chinese multi-voice text-to-speech (TTS) based on the VITS model that outputs WAV files; and third, a "push-to-talk" function in the WebUI that automatically transcribes recordings and fills them into the input box. The plugin supports WAV encoding and decoding, as well as resampling from 8kHz to 16kHz. It also includes built-in model management tools that support resuming downloads and downloading from domestic mirrors.
Read Moredsh-plugin-voice-input: Local voice to draft in DSH Web input
`dsh-plugin-voice-input` is a DeepSeek Harness Web voice input plugin developed by tangbut1, open-sourced under the MIT license. This plugin aims to address the inconvenience of inputting long sentences or mixed Chinese and English content in web input boxes. By adding a "hold to speak" microphone button next to the input box, it enables local offline speech recognition. Users press and hold the button to record audio, and upon release, the system performs offline recognition using the `sherpa-onnx-node` and `SenseVoice-SSmall INT8` models via the local loopback address (127.0.0.1:18765). It supports automatic language detection for multiple languages, including Chinese, English, Japanese, Korean, and Cantonese. The recognized results are appended to the current draft via DSH's official slots and events, and are not sent automatically.
Read More