audio.cpp 0.3 adds five local TTS families and extreme generation speed
Why fast local speech inference matters when it changes the workflows a team can actually own.
The interesting number in audio.cpp 0.3 is not ten hours of speech in three minutes.
It is what a number like that could make practical.
The project has added five TTS families, IndexTTS2, Irodori TTS, MOSS TTS Nano, MOSS TTS Local, and Supertonic 3, along with Chatterbox voice conversion. Its README reports that Supertonic 3 can generate about ten hours of audio in three minutes on an RTX 5090, and run at more than six times real time on CPU.
Those are project-reported benchmarks, not a reason to declare hosted TTS obsolete. Hardware, model choice, audio quality, concurrency, and the surrounding pipeline still decide whether a result holds for a real product.
But they do change the question.
Local TTS used to be easy to dismiss as a privacy-driven compromise: useful when data could not leave the building, too slow or awkward for much else. If a local runtime can generate long-form speech quickly enough, it becomes interesting for different reasons.
A team can test narration at scale without waiting on an API queue. An internal tool can generate drafts without sending sensitive material to another vendor. A product can batch audiobooks, onboarding flows, redubs, or accessibility audio where the economics of every generated minute are visible. Voice conversion can sit beside synthesis instead of becoming another hosted integration.
That does not make self-hosting free. It moves the bill into hardware, deployment, observability, model updates, and someone owning the bad day when inference slows down.
The benchmark would not settle it for me. I would want to hear the output on the material the product will actually use, then run it on the hardware expected to carry the load. Long inputs and concurrent jobs matter more than one clean demo.
The reason for owning the runtime matters too. Privacy or a large batch workload might justify the effort. If the only benefit is avoiding an API bill, the hardware and maintenance costs may wipe out the savings. Someone still has to own deployments, model updates, and the job that stalls at 2 a.m.
The value of audio.cpp 0.3 is not a leaderboard claim. It is another sign that local speech is becoming an infrastructure choice. For the right workflow, the trade-off may no longer be “hosted and usable” versus “local and slow.” It may be a more interesting choice between hosted convenience and local control.