QVAC is the fastest engine for audio on your own device

Our engineering team is continuously optimising the QVAC audio stack.

The optimisations live in qvac-fabric-speech.cpp, the open-source engine underneath it, and we benchmarked its three main capabilities: transcription, speech synthesis and music generation.

github.com/tetherto/qvac-fabric-speech.cpp

We compared them on four machines and six GPU lanes, against the open-source engines people run for each job. Here are the results.

Transcription

We transcribed a 98-second clip with Parakeet TDT 0.6b v3.

On a fanless MacBook Air it takes us 756 ms and audio.cpp 865 ms. That machine has no fan and its results swing 7 to 24 percent between runs, so we read those two as level. transcribe.cpp takes 1,746 ms on the same laptop. parakeet.cpp, the engine LocalAI ships for this model, has no working Metal path at the commit LocalAI pins, and takes 8.2 s.

The gap opens up on the other five GPU lanes, where Fabric leads audio.cpp by 1.85x to 2.81x. It leads transcribe.cpp on all six, by 2.1x to 8.5x. On an RTX 5090 over CUDA the clip takes 58 ms.

Speech synthesis

Supertonic 3 speaks 26.5 seconds of text in 0.65 s on that MacBook Air, against 4.70 s for audio.cpp.

Fabric finishes synthesis first on Metal, CUDA and Vulkan, 2.4x to 7.2x faster end to end than audio.cpp, and loads its model 2.3x to 4.8x faster.

Music generation

Generating a song takes us 9.0 s on the MacBook against 10.0 s for acestep.cpp, and 9.7 s against 10.0 s once our model load is counted in.

On a Strix Halo we generate the same song in 2.3 s against 4.8 s for acestep.cpp, and we come out faster on every one of the 60 prompt-and-lane combinations.

How we measured

• Transcription: the clip run once, then eleven times in the same process, and the difference taken, so the loading time drops out.
• Speech synthesis: the whole process timed, model loading included.
• Music generation: generation timed, both engines loading the same four GGUF files.

On transcription every engine ran at its own shipping defaults with its own weights. On synthesis both engines ran at f16, the only precision their two packages share. We quote no engine's own timer, including our own.

Versions tested: audio.cpp 0.7.3, transcribe.cpp 0.2.3, parakeet.cpp at the commit LocalAI pins, acestep.cpp at upstream HEAD. We forced acestep.cpp's release build, since its own script leaves the build type unset.

Try out QVAC Fabric Speech now: https://github.com/tetherto/qvac-fabric-speech.cpp