dsh-minimax-asr

Voice has been a missing layer in DeepSeek Harness: asking a model to understand an audio clip, hearing a spoken summary after a task finishes without staring at a screen, or simply conversing with an agent hands-free were all things you had to build yourself. dsh-minimax-asr bridges MiniMax asr-1.0 speech recognition and speech-2.8-hd synthesis into a single Harness global plugin, covering the full listen, transcribe, speak, and announce pipeline. It registers a transcribe_audio tool for the model that accepts wav, mp3, flac, opus, ogg and more, with output in plain text, verbose_json (speaker labels plus word-level timestamps), srt, or vtt, capped at 500 seconds and 50 MB per file. In the browser, a microphone button on the input toolbar records, transcribes, and appends the result to your draft in two clicks. After each task turn, the model calls announce_speech to write a concise spoken brief, the host pushes it over SSE, and the browser synthesizes it aloud; if no brief is written, only the first sentence of the reply is spoken, never the full report.

The most complete hands-free mode is the conversation button: tap it once and the microphone opens automatically. An adaptive energy gate (three times the measured noise floor, clamped to a 0.0025 absolute minimum and a 0.03 maximum) detects when you stop talking; 1.2 seconds of silence ends a turn, the clip is transcribed and submitted for you. While the agent thinks the microphone stays closed so it does not record its own speaker. The moment the spoken reply finishes the microphone reopens for your next sentence, and the turn counter is displayed on the button. A 30-second idle timeout releases the mic and pauses the loop so no credits are wasted. For voices, the host pulls the full 303 system voice list from MiniMax, grouped by language (Mandarin, Cantonese, English, Japanese, Korean, and more), with a ten-minute cache and an audition button so you can try any voice before saving. Credentials are resolved per-request through the credentials seam, defaulting to MINIMAX_API_KEY, whether stored in the local credentials file or exported as an environment variable.

This plugin is for developers who build with DeepSeek Harness daily and want voice as a first-class input channel. If you frequently feed meeting recordings, podcast clips, or dictated ideas into an agent, or you want to kick off a task by voice on a walk and hear the conclusion spoken back, this is built for that workflow. It also suits teams that want to evaluate MiniMax ASR and TTS endpoints without standing up their own service: four local routes (transcribe, speak, voices, events) plus a live diagnostics page expose wiring state, real-time energy gate values, browser listener counts, and the full microphone pipeline at a glance.