Create and Edit a Post Just by Talking | Rifff

The short answer

Rifff lets you create and edit content by speaking. Your words are transcribed in real time into text, and that text is what you send — but the text is not just a caption. It is instruction to an assistant that is building the piece while you talk, on a live canvas you can see. Hold the create button, talk, release; or tap it and type. The loop it is built around is say → watch it change → riff it into the feed. Follow, like, comment and message live in the same app. It requires iOS 17 or later, it needs both the microphone and speech recognition permissions, and it is documented as still in development. Here is what that actually means in practice. Rifff is here.

Why voice is the real question behind the query

Typing is an excellent way to write something whose wording you have already decided: an announcement, a caption, a careful reply. It is a poor way to capture something while it is happening — and content in a live feed is usually that.

Existing apps only half-solve this. Dictation into a composer is fast, but you still end up editing a wall of text before it is yours. Voice messages solve the opposite half: they capture the moment but leave nothing you can build on, edit or remix. The query behind "post by voice" is really "which half am I getting, and does the other half exist?"

Rifff takes the second half too. Voice is the input to a conversation with an assistant that can actually build. That makes the hard parts of voice the hard parts of the product, so they are worth spelling out before you decide whether it suits you.

The hold-to-talk mechanic, precisely

The create control has two behaviours, both advertised on the control itself: tap to open and type, hold to talk.

While you hold, words appear in the overlay as they are recognised. The text builds while you are still speaking — partial results are replaced as recognition improves, so the line on screen can change under you. Release sends it; swipe up while holding to cancel. Swiping into the cancel zone changes the hint to a cancel prompt, and releasing there discards the turn.

Two permissions are requested the first time, in order: Speech Recognition first, then the microphone, because neither is useful without the other. The overlay distinguishes the states that actually happen on a phone:

• Opening the microphone — permission is in flight; recording has not started.
• Listening — the microphone is live and the transcript is on screen.
• Didn't catch that — the recognizer produced nothing. Nothing is sent. The overlay tells you to hold and try again.
• Voice is off — permission was denied. The overlay names the exact place to fix it: Settings > Rifff, Microphone and Speech Recognition.
• Voice isn't available on this device — for example the simulator has no usable microphone. The overlay says so and closes, and the typed composer is still there.

The recognizer runs on the device's current locale and falls back to Simplified Chinese and then US English. If your language is not one it can handle, the fallback may be worse than nothing — that is a real limit, not a footnote. It also uses iOS's built-in speech recognition, so Apple decides whether processing is on-device or through its servers. Whatever you say on a network is audio that passed Apple's recognizer.

What your words actually do

This is the part a typical voice-to-text page will not tell you. The transcript is a turn sent to an assistant with real tools: reading and writing files, editing and deleting, finding files by pattern, searching file contents, and searching the web with quoted excerpts and cited sources shown in the transcript. So the loop is conversational, not one dictate-and-post. If your spoken instruction is ambiguous, the assistant can stop and ask you a question and wait for your reply.

The work in progress is visible, not summarised. The transcript shows thinking, each tool call with its argument and how long it took, search results with what was cited, and plain chat; a failed call expands to show its error. A phase indicator distinguishes thinking, reading, modifying and verifying, and the canvas shows the draft being changed as it happens.

Two mechanics behind it are worth knowing:

• One creation at a time. The session is a single global slot. Start a second while one is in progress and you are asked whether to end the current one or go back to it.
• Leaving the screen does not stop it. The session lifecycle is owned outside the page, so the work keeps running in the background and the transcript is replayed from disk when you return.

From a spoken turn to a published post

A creation starts as a draft in its own folder on your device, and the conversation is stored alongside it so the thread survives relaunch. The canvas is a live preview of the draft, with a badge while it is still a preview. When you are satisfied, the publish panel asks for a title, offers a category (personality film, quiz, generative, toy — or skip for now), and takes a palette; a remix inherits its parent's palette.

Publishing recompiles the draft, uploads a preview image, and only then marks it published. If any step fails, the draft is kept and the error is shown so you can retry. Three specific behaviours are worth knowing in advance:

• Publishing cannot be interrupted once it starts, though you can wait for it to finish.
• If the preview image has not rendered yet, publishing fails immediately and tells you to try again once it appears.
• A publish already in progress blocks starting a new one, with an alert rather than a silent failure.

A failed tool call is also never swallowed: the card expands to the underlying error, so you can see whether the assistant failed, was denied permission, or could not tell whether the effect happened.

What Rifff does differently

• The camera is never opened. The app declares a camera permission because the recording backend shares the system capture framework with it, and its permission copy says exactly that. If a camera permission has scared you off a voice app, this one tells you why it is there and admits it will not use it.
• No audio is stored. Voice is a way of talking to an assistant that builds your post, not a recording you keep.
• Memory across creations. The workspace holds a memory file of your preferences that carries from one creation to the next.
• A single persistent session, Apple Sign-In or email sign-in, and a dark-only visual treatment.

Honest limits: Rifff's own documentation lists it as in development, creation is iOS 17 and later, one creation runs at a time, publishing waits on a rendered preview, and a silent recording produces nothing. And the wait is real. A turn is cloud model inference plus sometimes a web search and a build check, which takes seconds, not the instant a dictation app finishes. If you expect to say a sentence and have a post appear, that is the gap to know about.

Who this is for

• For you if your ideas arrive faster than your typing and you would rather talk a rough instruction than write a polished caption, and especially if you riff something you just saw and want to talk through your version of it straight away.
• Not for you if you want a caption-ready post with no editing, if you want to publish several things in quick succession, or if you want your recording stored and replayed as audio. Rifff is for people who want to make something by saying what they want, not people who want to keep a recording.

Either way the decision is now easy: try holding the button for one sentence. If the transcript appears word by word and the assistant starts building, that is the app. If you would rather type, tap it — nothing is voice-only.