1. Follow the current Windsurf documentation redirect
The former Windsurf Cascade documentation URL now redirects to the current Devin Desktop Cascade page. That page describes Cascade as the agentic assistant in Devin Desktop and documents Voice input as one of its features.
Use the current documentation rather than an older screenshot or tutorial because the product name, setup, controls, and availability can move during this transition. One guide owns both Windsurf and Devin Desktop search intent instead of publishing duplicate pages for the same evolving product.
2. Use native Voice input when Cascade is the destination
Open Cascade with Command-L or the current Cascade control, choose the intended Code or Chat conversation, then use the visible Voice input control. The official page says the feature can transcribe your speech to text.
Native voice is the most direct option when you are already speaking to Cascade. It keeps the transcript attached to the active conversation and avoids adding another input layer for that one surface.
- Confirm the repository and active Cascade conversation.
- Speak one bounded requirement.
- Review the transcript before the agent receives ambiguous names, paths, commands, numbers, or negations.
3. Keep Cascade voice separate from editor-wide text input
The official Cascade page scopes Voice input to interacting with Cascade. It does not describe that control as a voice keyboard for the code editor, integrated terminal, commit-message field, browser preview, or other applications.
For another Mac text field, click until the insertion point is visible and use the Dictation shortcut shown under System Settings → Keyboard. Apple says Dictation can enter text anywhere you can type, subject to the current field, language, Voice Control, microphone, and processing configuration.
4. Do not invent a privacy claim the voice page does not make
The current Cascade overview says Voice input transcribes speech to text, but the reviewed page does not identify the speech processor, transmission path, retention rule, or a local-only mode for that audio. Do not label it local or cloud from the interface alone.
Keep transcription and agent processing separate. Even if one speech stage were local, the resulting prompt, selected text, repository context, tool calls, and model work would still follow the active Devin Desktop account, model, and service controls.
5. Review speech before Cascade can act on it
A voice transcript in an agent prompt can become more than prose. Cascade Code mode can edit files and call tools, while Chat mode can propose code. A misheard path, package, flag, permission, or destructive verb can materially change the task.
Use a visible review step before submission. If the request is long, speak the goal, constraints, acceptance criteria, and exclusions as separate points rather than one unstructured paragraph.
6. Add IraVoice for local dictation and code names
IraVoice 1.2.2 is a separate option on Mac, Windows, and Linux when one held-key workflow should produce editable text across developer, email, document, and browser fields. On IraVoice Local, speech recognition runs on your computer with NVIDIA Parakeet after a one-time model download.
In the Windsurf editor and terminal, IraVoice finds the Git project you are working in and spells its code names, so “use effect” becomes useEffect; it reads only names, never source text, and nothing leaves your computer. IraVoice does not control Cascade, dictate reliable source syntax, press Send, or bypass the destination’s account and tool permissions.