Feature Request: Voice Mode

[NichUK]

I'm keen to be able to chat with my Claw, so I've started to build Voice Mode, as an add on to the current tray app.
To make it flexible for everyone, I'm building two separate modes, AlwaysOn, and Wakeword. I'm also allowing for Windows built in STT and TTS as well as user-configured STT/TTS via configuration (e.g. Minimax, ElevenLabs, etc), and integration with the current chat function, rather than just an invisible background function (although it will be an invisible function if Chat is not open at the time).

@shanselman I don't know if this is already on your roadmap, but would you be open to a PR? I already have AlwaysOn with Windows STT & TTS integrated with Chat working well, and I have deliberately architected it so as to fit with the rest of the project, but with minimum touchpoints (can also probably do a little more work there).

With your permission I'll generate a PR for review and discussion, and you can let me know if the architectural approach is good for you. I'm also in the process of adding the additional TTS providers as the Windows TTS is a bit old-school. I've also done a design document which lays it all out.

Nich

[shanselman]

Sounds like fun and a great idea @NichUK! I would ask that you "light up with the best of windows" - there is the old TTS stuff and there's newer models and APIs. I think there are 3 ways to do speech on windows that is built in, and then there's new stuff like http://Handy.computer and whisper, so as long as it's architected in a way that let's us grow I'm into it! I can also have the windows speech team look at it as well! :)

[NichUK]

I've got stuff working well, but just making it a bit nicer before I share and reducing the latency as much as possible - although the biggest latency is OC response... Should have a PR tomorrow.

[github-actions]

🤖 This is an automated response from Repo Assist.

Great work, @NichUK! This looks like a natural fit for the existing capability plugin architecture — a few pointers that might be useful:

Where voice plugs in

The INodeCapability / RegisterCapability pattern in WindowsNodeClient is the right integration point. Each capability handles a category of commands dispatched from the gateway. Existing capabilities (CameraCapability, ScreenCapability, SystemCapability) in src/OpenClaw.Shared/Capabilities/ show the pattern — you implement INodeCapability, wire events to the UI layer, and register in NodeService.cs (~line 108).

Windows speech API landscape

As @shanselman noted, there are a few tiers to consider:

Tier	API	Notes
Modern WinRT (preferred)	Windows.Media.SpeechSynthesis + Windows.Media.SpeechRecognition	Available in-box on Win10+; supports the newer neural voices
Legacy .NET	System.Speech	Broad compatibility but older SAPI5 voices
External / premium	Azure Cognitive Services Speech SDK, Whisper (local), ElevenLabs	Excellent quality; requires opt-in config

Windows.Media.SpeechSynthesis gives access to the same neural voices that power Narrator and modern Edge — much more natural than SAPI5 and zero extra dependencies.

Suggested architecture sketch

VoiceCapability : INodeCapability
  ├── ISTTProvider (interface)   ← WinRT / System.Speech / Whisper impl
  └── ITTSProvider (interface)   ← WinRT / Azure / ElevenLabs impl

This matches the extensible approach @shanselman described and keeps the plugin self-contained. Looking forward to the PR!

Generated by Repo Assist · ◷

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@cbb46ab386962aa371045839fc9998ee4e97ca64