voice-call

Voice call · 语音通话

Speak in your current DSH conversation. Hear your Agent reply.

Start the microphone, speak, and pause: the plugin transcribes and sends your message, then reads the response aloud. Keep using Standard mode and your conversation’s existing model, tools, and permissions. No separate Agent preset is needed.

中文指南 · npm · GitHub · Releases

Actual usage screenshot, with the entire sidebar removed to exclude workspace names, conversation lists, and account details. The model shown is an example, not a requirement. Screenshots use the Chinese interface.

Actual connected call view. The text editor is hidden and the bars respond to microphone volume. Workspace and branch names above the call panel have been cropped out.

What you can do

Get started in three steps

1. Install

Requires DSH 0.2.2-rc.1 or later in the 0.2.x line, a microphone, and working audio output.

In DSH, open Plugins → Add plugin and add this npm package:

voice-call

Check that the source repository is crease123/voice-call to distinguish it from similarly named projects. Once installed and enabled, the composer should show a microphone, and Settings should include “Voice call” and “Voice call instructions.”

Alternatively, get the prebuilt package from GitHub Releases and add its download URL in DSH. There is no need to download or build the source.

dsh plugin --profile <your-profile> add voice-call

This guide describes v0.3.4 and later, where voice calls work within your existing conversation without a separate Agent preset. Upgrade older installations to use the behavior shown here.

2. Configure your speech services

Open Settings → Voice call. Select recognition and synthesis services independently. Recognition turns your speech into text; synthesis turns the Agent’s reply into audio. These choices do not replace your conversation’s chat model.

Either Doubao credential method works; API key takes priority. An unconfigured API key does not prevent legacy authentication when both App ID and Access token are configured.

Saved secrets are not displayed again; blank fields preserve existing values. Credentials stay in your DSH Host credential store. The plugin supplies no shared key, author account, or free quota. Your provider account covers speech-service usage charges.

Apply your selections before previewing speech. After switching providers, hang up and start a new call. DeepSeek offers Stella and Tide voices; rate, pitch, and delivery controls apply to Doubao only.

3. Tap the microphone and talk

Open an ordinary conversation and click the microphone at the lower right of the composer. On first use, allow microphone access for DSH or your browser. Start speaking when the status says “Listening.”

Use the microphone marked by the arrow, rather than the Send button to its right.

Pause after a sentence to transcribe and send it automatically. Hear the Agent’s response while retaining the conversation’s text transcript. Speak to interrupt playback when you need to change direction. Click Hang up voice call to return to text input; your unsent text draft is preserved.

Make it work for you

Voice, volume, and capture

Use Settings → Voice call to adjust:

• Playback volume: affects speech in this client only, without changing system volume.
• Interruption sensitivity: lower it if background noise interrupts; raise it if quiet speech is missed.
• Minimum speech duration: increase it to filter brief noises. Too high a setting may miss very short utterances.
• Preview: check the selected voice. Preview temporarily pauses capture and automatic readout in active calls, then resumes them afterward.

Headphones help prevent the microphone from picking up speaker playback.

Instructions that apply only during calls

Open Settings → Voice call instructions, edit the rules, and save. For example:

Answer the main question first, usually in 1–3 natural sentences. Leave code and long lists in text; speak only the key points.

Saved rules apply to the next model request, without rewriting a response already in progress. They apply to any connected voice conversation and are withdrawn on hangup or disconnect. They do not replace the existing persona, tools, or permissions. The plugin no longer adds a separate Personal Agent / Voice call preset.

Troubleshooting

Recognition and synthesis depend on network and provider latency. DeepSeek recognition uploads a completed audio segment after each utterance; continuous conversation does not imply word-by-word live transcription.

Development and architecture

Users only need to install voice-call. The main release embeds its three runtime components.

Source builds require a matching Harness checkout. First run pnpm run build:artifacts there, then run in this repository:

pnpm install
DSH_HARNESS_ROOT=/absolute/path/to/deepseek-harness pnpm run build

For local watching, use pnpm run dev:watch with the same DSH_HARNESS_ROOT. Configure source entries in the profile’s own cordis.patch.yml so plugin settings can be persisted. Remote generation temporarily changes Harness build configuration; do not run it alongside other edits to that configuration.

Read more: Speech subsystem · Call context · Bundle.

MIT License. Never commit API keys or tokens, or include them in issues or screenshots.

Version 0.3.4 preserves speech playback ownership when runtime context is injected, fixing silent first replies after starting a call.

Continued speech while recognition is pending is retained and merged in order before submission; speaking during an Agent reply still interrupts playback.