dsh-voice-by-edge

Chatting with a model in the DSH web app usually means staring at a wall of text. dsh-voice-by-edge turns that silent reading loop into a listening-and-speaking loop: flip on the Voice Practice toggle and the model starts replying in short, spoken-style sentences that the browser reads aloud immediately, while a microphone button lets you dictate your next turn by voice.

The TTS layer uses Microsoft Edge neural voices (edge-tts) by default—Aria for English, Xiaoxiao for Chinese—splitting mixed-language segments automatically so each line gets a matching voice. If the Edge endpoint is unreachable it falls back to the browser's built-in speechSynthesis, so you never lose audio. Speaking speed is adjustable from 0.75× to 1.5×. Speech input rides on the Web Speech API: recognized text appears in the composer in real time and can auto-send. On the server side the plugin injects a voice-practice prompt into the session's system instructions, capping reply length, stripping out tables and code blocks, and appending a gentle correction when your phrasing slips, keeping every turn in a comfortably speakable register.

It is built for DSH web users who are prepping for a speaking exam, want to drill daily listening and pronunciation, or simply believe a conversation with an AI should also have a voice. Switch the mode on and start practicing; switch it off and you are back to plain text. No extra client software, no persistent state beyond the session.