dsh-doubao-voice

Typing into DeepSeek Harness for long stretches gets tiring. dsh-doubao-voice removes that friction: hit a global hotkey (default Alt+V), speak, and transcribed text streams into your input field word by word. You review, tweak, then send. Under the hood it rides Volcengine Doubao Seed ASR true streaming recognition, with 16 kHz PCM sent in 200 ms frames, so tokens appear while you are still talking rather than arriving in a batch after you stop.

A few details make the day-to-day experience solid. A four-state visual indicator keeps you oriented through preparing, listening, and wrapping up. An LCP merge algorithm ensures mid-utterance corrections and long recordings with multiple pauses never pile up as duplicate text in the field. Transcripts are never auto-sent; they land in whichever editable element holds focus, with a clipboard fallback if the write fails, keeping the pipeline resilient. The settings panel lets you manage Volcengine credentials, up to 100 hot words, target language (zh-CN, en-US, ja-JP, and other BCP-47 tags), and maximum recording duration. Credentials are stored locally with strict file permissions scoped to the current user.

Built for DeepSeek Harness power users and content creators who live in voice input, especially those who switch between Chinese, English, and Japanese and want a transparent pipeline they can edit, roll back, and fine-tune with custom vocabulary. Community project under the MIT license; please review Volcengine service terms before use.