kittcat-lab/dsh-kitt-voice
DSH KITT adds a spoken conversation to the DeepSeek Harness web interface, built Spanish-first: a Silero turn detector decides when you have finished speaking, the reply is read aloud sentence by sentence as it arrives, you can interrupt by talking, a companion Electron window floats above other apps with global keys, and local Piper voices can keep synthesis entirely offline.
DSH KITT adds a spoken conversation to the DeepSeek Harness web interface, built Spanish-first and usable without the browser in front of you. Press a key and talk: a Silero v5 turn detector (not a loudness meter - measured thresholds separate silence 0.04 from engine noise 0.13) decides when you have finished, the text is transcribed and sent, and the reply is read aloud sentence by sentence as it arrives, with a per-reply echo baseline so talking over the reply interrupts it and headphones adapt automatically. It offers push-to-talk when you prefer it, one voice per reply (falling back to the built-in voice and saying so if the engine fails), 104 neural voices (45 Spanish, 47 English, 12 Chinese) plus local Piper and the system voice, adjustable speed, microphone/output chosen by name, and a bilingual English-word pronunciation list for Piper. A companion Electron window floats above fullscreen apps with its own controls, a state-carrying border, global keys and a menu; and it checks its caller on all thirteen loopback routes, keeps the key out of the browser, and refuses to let a voice or file name become a path. 115 tests ship with it.
Install
dsh plugin --profile web add dsh-kitt-voiceCopied verbatim from the README's Install block; the README notes the command appends the plugin to the profile's bundle list by itself, and documents a link: local-checkout form and removal with dsh-kitt-voice.
Compatibility
The README states it has only ever been run on Windows. The default browser recogniser works in Chrome and Edge but not inside Electron (the plugin switches to Groq Whisper and says so); the transcription key comes from the harness's own credential store (GROQ_API_KEY, or DSH_KITT_API_KEY / GROQ_API_KEY in the environment) and never reaches the browser. Neural voices read the reply through Microsoft's read-aloud service - the README states plainly that the text of the reply leaves the machine - while local Piper voices and the system voice are offline options. English, Spanish and Simplified Chinese READMEs.
Details
- Repo: kittcat-lab/dsh-kitt-voice
- Category: Vision & Multimodal
- Stars: 0
- Version: GitHub repo kittcat-lab/dsh-kitt-voice (0★, MIT); npm package dsh-kitt-voice at 1.3.0 (registry check 2026-10-13; the registry repository field resolves back to this repo).
Alternatives
54xkeee/dsh-vision · 54xkeee/dsh-youreyes · Cheng-cheng9669/dsh-deepseek-vision