[audio.cpp] Release 0.7: 62 audio model families (85+ variants), Arena UI for model comparison, MiniMax Music 3, FireRed TTS3/Audio, ControlFoley, Personaplex, and more

audio.cpp 0.7 is out :)

This release adds a lot of new audio models and a new way to compare them locally.

Audio.cpp is now at 62 model families and 85+ model variants. And it keeps growing!

The biggest user-facing change is the new Arena UI. Instead of testing one model at a time, you can now give one shared input and queue multiple local models or GGUF variants, then compare the generated outputs side by side. This is useful for picking between models without writing a pile of scripts.

Disclaimer: the RTF numbers are from cold one-shot requests using the current audio.cpp implementations (+server overhead), so don’t use them as a model leaderboard. If one model is slower, it might just mean my implementation still needs optimization. The goal is to help you try a bunch of models locally, compare the outputs, and pick the one you like best.

Expanded In 0.7

• TTS / Voice: FireRedTTS3, MagpieTTS, PersonaPlex, F5-TTS / Habibi, MOSS VoiceGenerator, DotTTS Edit
• ASR / Speech Understanding: FireRedAudio, IBM Granite Speech 5.0 TurboCTC, MMS Forced Aligner
• Voice Conversion: MeanVC2
• Music / Audio Generation: MiniMax Music 3, MiDashengLM-Gen, ControlFoley (experimental), ACE-Step 1.5 XL variants
• Audio Tools: AudioSR

A lot of the new coverage happened because contributors helped bring models up quickly, sometimes very close to day one after release!

What I’m most excited about is seeing audio.cpp run well on real edge hardware: Our contributor https://github.com/Hi5808 tests audio.cpp on NVIDIA Jetson Orin: 40/40 model families works without issues on Orin NX 16GB and 34/40 on Orin Nano 8GB.

Our prebuilts now cover Windows CPU, Windows Vulkan, Windows CUDA 12.4, Windows CUDA 13.3, Ubuntu x64 CPU, Ubuntu x64 Vulkan, macOS arm64 Metal, macOS x64 CPU. Thanks https://github.com/drzsdrtfg for adding the automated prebuilt workflows and freeing me from manual release builds.

Finally, contributions are very welcome! If you are interested in local audio AI, model integration, performance, deployment, UI, or just testing things on your own hardware, I’d love to have you involved.