Oreja de Van Gogh wants to pause Apple Music when ChatGPT starts talking

Oreja de Van Gogh is a Mac utility with a one-line pitch: it pauses Apple Music while ChatGPT speaks. That is the whole product. No model, no agent framework, no API. A listener is playing music, a voice assistant starts talking, and the utility decides the music should yield. The listing on Product Hunt carries the flat description “pause Apple Music while ChatGPT speaks,” and that is enough to understand the thing.

It is also, read correctly, a complaint. The fact that someone had to build this means macOS does not arbitrate audio between a music player and a conversational AI. Every other piece of the stack has been built out: speech-to-text, the language model, text-to-speech, the wake word. What has not been built is the boring middle: a system-level rule for who owns the speakers when two applications both want them.

The gap is arbitration, not intelligence

Voice assistants have shipped for a decade, and the failure mode has not changed. The assistant talks over whatever is playing. Apple’s own Siri ducks music on iPhone and HomePod, because Apple controls both sides of that transaction: the OS, the music app, and the assistant. ChatGPT on the Mac is a third-party application. Apple Music is Apple’s. The operating system has no obligation to mediate between them, and evidently does not.

So the workaround is a small app that watches for one process and sends a pause command to another. That is a hack, and it is the kind of hack that tells you where the platform is thin. Compare it to the printer dialog, or the camera permission prompt, or the microphone indicator in the menu bar. Those exist because the OS decided that shared resources need a referee. Audio output has a referee for volume and device routing. It does not have one for conversational turn-taking.

This matters more as voice becomes a primary interface. OpenAI has pushed hard on real-time voice since the GPT-4o launch in May 2024, and the company’s Advanced Voice Mode is designed to be interruptible, low-latency, and conversational. That design assumes the assistant is the only thing making sound. In practice, users listen to music, take calls, run a podcast, join a meeting. The assistant is one voice among many, and nothing in the stack knows that.

What a real audio layer would look like

The interesting question is not whether Oreja de Van Gogh works. The interesting question is why the problem is being solved at the application layer at all.

A proper solution lives in the OS. It would define a priority class for “conversational agent” audio, distinct from media playback, distinct from notification sounds, distinct from a phone call. When an agent in that class starts speaking, the OS would duck or pause lower-priority streams, then restore them. Android has partial versions of this with audio focus requests, and iOS has audio session categories that developers can set. Neither is designed for a world where a third-party AI assistant is a persistent conversational presence on a desktop.

Microsoft has a version of this problem too. Copilot on Windows can speak, and Windows has a long history of audio session management through WASAPI, but the arbitration is still per-application and coarse. The result is the same: users patch the seams themselves.

The commercial implication is that the assistant vendors have an incentive to solve this, and the OS vendors have an incentive to control it. If Apple builds a first-class “agent audio” session type into macOS and iOS, it owns the arbitration and can privilege Siri. If OpenAI builds its own arbitration layer, it can make ChatGPT the polite citizen that ducks Spotify, Apple Music, and everything else, and in doing so make itself the default voice on the machine. Oreja de Van Gogh is a hint that neither has done the work yet.

The economy of tiny fixes

There is a business story here too, and it is smaller and more honest than the AI-hype version. Product Hunt is full of utilities like this: single-purpose Mac apps that fix a specific friction for a specific user. They rarely become companies. They sometimes become features. The ones that matter are the ones that get absorbed: the OS vendor ships the fix, or the assistant vendor ships the fix, and the utility evaporates.

That is the likely fate of Oreja de Van Gogh, and it is not a bad fate. The utility is a probe. It tests whether users care enough about the friction to install something. If they do, the platform vendors will notice, because the friction is now measurable.

The harder question is what the probe reveals about the AI economy more broadly. A lot of the value in the current wave is being captured at the model layer, where the labs compete on benchmarks and price per token. A lot of the friction is being created at the integration layer, where the model meets the operating system, the browser, the music player, the calendar. The labs are not optimized to fix integration friction, because it does not show up in an eval. The OS vendors are not optimized to fix it either, because the assistant is someone else’s product.

The most interesting AI products of the next two years may not be models. They may be the small pieces of glue that make models behave like they belong on the machine.

That is where the tiny utilities live. Not in the frontier, but in the seams. A pause button for Apple Music is a trivial thing. It is also a precise measurement of how far the assistant stack still has to travel before it feels native.

What to watch

Two things. First, whether Apple ships an audio session category for conversational agents in a future macOS or iOS release. The company has been slow to open up system-level hooks to third-party assistants, and the DMA in Europe is pushing it to open more. If Apple does ship it, the Oreja de Van Goghs of the world disappear, and that is the signal that the platform has caught up.

Second, whether OpenAI, Anthropic, or Google build their own arbitration into their desktop clients. The assistant that politely ducks your music is a better citizen than the one that talks over it, and being a better citizen is a competitive advantage on a machine where the user has other things going on. The first vendor to ship a clean “agent audio” mode will win a small but real amount of goodwill from exactly the users who run assistants all day.

Neither has shipped it yet. That is why a one-feature Mac app made it onto Product Hunt, and why the listing is worth reading as a status report rather than a product announcement.