DeepSeek Harness plugins directoryDSH Pluginsdsh plugins directory
Back to guides

DeepSeek Harness voice input and TTS: speak, then listen

Two short Bilibili recordings, eight screenshot-verified steps: tick local transcription in the DeepSeek Harness Web UI, dictate with Ctrl+Alt+V, and let the dsh-plugin-tts plugin read replies aloud in a voice you picked.

Last updated: 2026-09-23

DeepSeek Harness Web UI composer with the voice-input mic turned into a red recording button, dictation armed and waiting for speech.
Recording state: the mic slot becomes a red stop button in the composer toolbar.

Two scenarios, two recordings. In the first, a creator talks to the DeepSeek Harness Web UI and his speech turns into text without leaving the machine; in the second, another creator has replies read aloud through dsh-plugin-tts. Neither video carries subtitles, so every quoted interface string below was read off the frames, and each step carries the original screenshot.

Everything shown runs on the local Web UI (127.0.0.1:3080 in the recording). Voice input rides on a local transcription plugin — dsh-voice-local by Real-WangLe — while playback rides on dsh-plugin-tts by 1624318455; the voice collection lists both lines and more.

The short version

  • ▸Voice input is one checkbox: 语音输入（本地转写，音频不出本机） — transcription runs on your machine, audio never leaves it, and no API key is involved.
  • ▸Ctrl+Alt+V is the default shortcut to start and stop dictation — custom bindings are supported, and words land in the composer while you speak.
  • ▸dsh-plugin-tts reads replies aloud with free Microsoft Edge TTS voices — speed, pitch and volume are sliders, and a pink speaking badge appears during playback.
dsh本地语音输入插件更新了还得是未来可期 · Bilibili
Watch on Bilibili
dsh-plugin-tts 让你的DSH能按你喜欢的样子开口说话我快不认识然字了 · Bilibili
Watch on Bilibili

Voice in, voice out, step by step

Talk to the AI — local voice input

  1. 1

    Open the composer's voice popover and tick local transcription

    In the DeepSeek Harness Web UI, click the mic slot on the composer toolbar. The popover exposes the voice-input switch — 语音输入（本地转写，音频不出本机）, local transcription with audio that never leaves this machine — plus 键盘输入自动关闭麦克风, which hands the mic back the moment you type. Both were ticked in the recording, and dictation keeps running even with the popover left open.

    DeepSeek Harness voice-input popover showing the local-transcription checkbox, the keyboard auto-release checkbox and the Ctrl+Alt+V shortcut row.
    The popover up close: local transcription ticked, keyboard auto-release on, shortcut listed.Watch at 0:34
  2. 2

    Press Ctrl+Alt+V and watch words land while you speak

    The same popover shows the shortcut row: Ctrl+Alt+V by default, and the presenter says custom bindings are supported. While recording, the mic slot turns into a red button; dictated text streams into the composer sentence by sentence, and a second press stops the mic, leaving the finished paragraph ready to send or edit.

    DeepSeek Harness composer mid-dictation: the red recording button stays active while transcribed Chinese paragraphs stream into the input box.
    Live dictation: text lands sentence by sentence while the mic is still red.Watch at 1:24
  3. 3

    Read the privacy fine print in the plugin README

    The featured repo, Real-WangLe/dsh-voice-local, states it plainly: recording and transcription both finish locally, with no cloud API and no API key; the SenseVoice engine (CPU inference) handles Chinese with automatic punctuation plus English, Japanese, Korean and Cantonese. Install is one line — dsh plugin --profile web add dsh-voice-local — and the SenseVoice model (~230MB) downloads on first mic click.

    GitHub README of dsh-voice-local stating that DeepSeek Harness recording and transcription finish fully offline with no cloud API, plus the one-line install command.
    The plugin README: fully local, no API key, SenseVoice engine, one-line install.Watch at 2:02
  4. 4

    Know the open Windows bug before you blame your mic

    Issue #1 on the repo — still open when the video was recorded — documents a Windows-specific defect on v0.3.0: after Silero VAD's gate misfires, the internal speechMs counter stays at 0 and /transcribe keeps returning empty text, even though the model reports ready. The reporter's debug evidence rules out volume and threshold settings, so if Windows dictation comes back empty, check this issue first.

    GitHub issue #1 on dsh-voice-local reporting that Silero VAD returns speechMs=0 on v0.3.0 under Windows, leaving DeepSeek Harness transcription permanently empty.
    Open issue at recording time: v0.3.0 on Windows returns empty transcripts.Watch at 1:50

Let the AI talk — TTS playback

  1. 5

    Pick a free Edge TTS voice in dsh-plugin-tts settings

    In the Web UI settings under 插件 → 语音, the TTS provider dropdown offers Edge TTS: free online voices, no key, working out of the box. The recording picks 晓晓 (zh-CN-XiaoxiaoNeural); speed, pitch and volume are −50% to +50% sliders, and 运行诊断 one-click checks online synthesis, the local RVC service and loaded models.

    DeepSeek Harness settings dialog for dsh-plugin-tts with the Edge TTS provider, the Xiaoxiao voice dropdown and speed, pitch and volume sliders.
    TTS provider set to Edge TTS with the free Xiaoxiao voice selected.Watch at 0:40
  2. 6

    Route through RVC if you want your own voice

    Switching the provider to 自定义音色（RVC） points the plugin at a local conversion service (default http://127.0.0.1:4892). Fill in your trained .pth model and an optional .index file, or pull a ready-made voice pack from a shared repository — the recording browses the author's own 1624318455/rvc-for-tts pack list.

    DeepSeek Harness dsh-plugin-tts dialog on the RVC route showing the local conversion service address, the source voice dropdown and .pth model path fields.
    Custom voice route: local RVC service, your own .pth model or a shared pack.Watch at 0:08
  3. 7

    Preview a line and catch the pink speaking badge

    Type a test line into 试听测试 — the recording uses 你好，这是一个语音测试。 — and hit play. While it speaks, a pink 发音中 (speaking) badge floats beside the controls; the dialog footer credits node-edge-tts as the engine.

    DeepSeek Harness TTS preview box with a test sentence, its play button and the pink speaking badge shown while dsh-plugin-tts reads the line aloud.
    Preview test: one line, one click, and the speaking badge lights up.Watch at 0:56
  4. 8

    Let replies read themselves aloud

    Send a message and the reply comes back with its own playback state: while the answer is being read, the message toolbar shows a stop control and the pink speaking badge follows along. The question in the recording — 你是谁 — gets a spoken self-introduction seconds after the text lands.

    DeepSeek Harness chat reply being read aloud by dsh-plugin-tts, with the stop control on the message toolbar and the pink speaking badge over the conversation.
    Live playback: the reply gets a stop button while Edge TTS reads it.Watch at 1:08

FAQ

Questions the frames raise but do not answer on their own.

Does voice input send my audio to the cloud?

No. The setting itself says 本地转写，音频不出本机 — local transcription, audio never leaves the machine — and the plugin README confirms recording and transcription both finish locally with no cloud API and no API key. The SenseVoice model (~230MB) downloads once and runs on CPU.

Which TTS engines can read replies aloud?

The recording uses Edge TTS: free online Microsoft voices such as Xiaoxiao, no API key, powered by node-edge-tts. For a custom voice, the provider dropdown switches to RVC, which runs a local conversion service against your own trained model — and a diagnostics button checks both paths.

Is the Windows empty-transcription bug fixed?

It was still open when the video was recorded: issue #1 reports that on v0.3.0 under Windows, Silero VAD's gate keeps speechMs at 0, so /transcribe returns empty text. Check the repo's issues page for the current status before filing a new report.

Can I change the Ctrl+Alt+V shortcut?

Yes. The popover lists Ctrl+Alt+V as the default binding, and the presenter says custom bindings are supported. One press starts the mic, a second press stops it.

Related guides

The rest of the voice track, plus the setup guides behind it.

The DeepSeek Harness voice collection

ASR input, TTS playback and duplex lines — every voice plugin curated with verified stats.

Read the guide

dsh-plugin-tts in the plugin directory

Repo facts, stars and install notes for the Edge TTS reader featured in this guide.

Read the guide

Install DeepSeek Harness plugins

The dsh plugin add workflow, profiles, and how to verify a bundle before it runs.

Read the guide

Run DSH on local models

Keep the whole loop on your machine — a natural pairing with local speech transcription.

Read the guide

Remote access to your DSH Web UI

Reach your voice-enabled setup safely when the machine with the mic is elsewhere.

Read the guide

Sources and credits

Both recordings come from Bilibili — two different creators, credited separately — and every screenshot deep-links to the exact second. Want more voices to choose from? Browse the voice collection

dsh本地语音输入插件更新了还得是未来可期 · BilibiliWatch on Bilibilidsh-plugin-tts 让你的DSH能按你喜欢的样子开口说话我快不认识然字了 · BilibiliWatch on Bilibili

DSH Plugins is an independent community directory of DeepSeek Harness plugins. Not affiliated with or endorsed by DeepSeek. Third-party plugins are not security-audited — review the source before installing.

New DeepSeek Harness plugins, weekly. No spam.