On July 14, 2026, the open-source project audio.cpp shipped its 0.3 release, adding five new text-to-speech families to a single C++ engine that runs entirely on ggml with no Python dependency. The headline number belongs to Supertonic 3: about 10 hours of audio generated in 3 minutes on an RTX 5090, which works out to more than 200 times faster than real time on CUDA and roughly 6 times faster than real time on a plain CPU. For creators who want studio-grade voice without a per-character cloud bill, that is a meaningful shift.
The update pushes the engine's model count past 30 families spanning text-to-speech, speech recognition, voice conversion, and music, all under a permissive Apache 2.0 license from ShugoAI LLC. This deep dive breaks down the four new voice models, explains why an on-device C++ runtime matters for creators, and walks through a five-step path to generating your first local voiceover.
What Shipped in audio.cpp 0.3
audio.cpp is a pure C++ inference framework for audio models. Version 0.3, dated July 14, 2026, adds Supertonic 3, IndexTTS2, Irodori-TTS, MOSS-TTS-Nano, and MOSS-TTS-Local, and extends the existing Chatterbox model with voice conversion so a single install now covers text-to-speech, voice cloning, and speech-to-speech. The release also broadens GGUF loading and conversion, so more of the model zoo can be quantized and shipped as portable single files.
Everything runs across CUDA, CPU, Vulkan, and Metal, and the project publishes prebuilt Windows binaries alongside Linux and macOS build scripts on its releases page. The team reports that its TTS paths run 1.8 to 5 times faster than the original Python reference implementations, with 45 to 80 percent lower latency. As a concrete example, a VibeVoice 1.5B model renders a 93.9-minute podcast in 18.2 minutes on their test rig. It follows the same C++ and ggml pattern as speech recognition ports like parakeet.cpp, but under one roof for the whole audio stack.
The Five New Voice Models, Compared
Each of the new voices targets a different job, from lightning-fast batch narration to precise dubbing and expressive character work. Here is how they line up.
| Model | Languages | Best for | Standout trait |
|---|---|---|---|
| Supertonic 3 | 31 languages | Fast batch and streaming voiceover | About 99M parameters, no GPU needed for fixed voice, 47 ms first-token latency |
| IndexTTS2 | Chinese, English | Precise timing and dubbing | First autoregressive zero-shot model with exact duration control |
| Irodori-TTS | Japanese | Expressive, character voices | Emoji-driven style control via flow matching |
| MOSS-TTS-Nano | Auto-detected | Lightweight real-time streaming | Small footprint tuned for low-latency streaming |
| MOSS-TTS-Local | Auto with optional hint | Long-form and multi-speaker | Stable long-form speech and multi-speaker dialogue |
Supertonic 3 is the standout for raw throughput. At roughly 99M parameters across its public ONNX assets, it is a fraction of the size of the 0.7B to 2B class open TTS systems, which is exactly why it can hit those speed figures on modest hardware.
Why Local TTS Is Suddenly Practical
Two years ago, running a competitive text-to-speech model locally meant a heavyweight Python stack, a capable GPU, and a lot of patience. audio.cpp 0.3 removes all three barriers at once. A pure C++ binary has no Python environment to manage, GGUF quantization shrinks weights into a single portable file, and small models like Supertonic 3 run comfortably on a CPU. That combination turns local voice from a research curiosity into a production tool.
The economics matter too. Cloud TTS services bill per character or per second of audio, which adds up fast for an audiobook, a series of YouTube videos, or a game with hundreds of dialogue lines. A local engine converts that recurring cost into a one-time hardware cost and keeps every script private on your own machine. For creators weighing quality trade-offs, the recent blind test of 46 voice models is a useful reference for which open voices hold up against paid ones.
Run Supertonic 3 Locally in Five Steps
Supertonic 3 is the easiest of the new models to try because its fixed voices need no GPU. Here is the general path from a clean machine to a finished WAV file. Check the repository for the exact command flags, which change between releases.
- Build for your hardware. Clone the repo and run the backend build script, for example
scripts/build_linux.sh --backend cudaon Linux with an NVIDIA GPU, or the Metal helper on Apple silicon. On Windows, download the prebuilt binary from the releases page instead. - Get the weights. Pull Supertonic 3 from its Hugging Face page and convert it to GGUF with the included conversion tooling, or grab a ready-made GGUF if one is published.
- Pick a fixed voice. Supertonic 3 ships fixed voices that run without a GPU, so select one for your first pass to keep the setup simple.
- Generate speech. Point the command-line tool at the GGUF model and a text file, and it writes a WAV you can drop straight into an editor.
- Scale up or stream. For long-form narration, batch a full script; for interactive apps, switch on CUDA streaming mode for roughly 47 ms first-token latency.
What This Enables for Creators
The practical payoff is a self-hosted voice studio that costs nothing per line. Audiobook narrators can render a full manuscript overnight on a single GPU. YouTube creators can generate placeholder or final voiceover without touching a subscription. Game developers can voice hundreds of NPC lines in-house, and accessibility teams can add narration to any document offline. Because Chatterbox now includes voice conversion alongside its cloning path, you can also record a rough take and reshape it into a cleaner delivery without re-recording.
Try it this week: build audio.cpp for your machine, generate a one-minute Supertonic 3 sample from a script you already have, and compare it against whatever cloud voice you currently pay for. If the quality clears your bar, every voiceover after that is free.
Frequently Asked Questions
Is audio.cpp free to use?
Yes. audio.cpp is open source under the Apache 2.0 license from ShugoAI LLC and runs entirely on your own hardware, so there are no per-character or per-second fees. Review the repository LICENSE file for redistribution and commercial-use terms.
Do I need a GPU to run these models?
No for Supertonic 3 fixed voices, which the project says run about 6 times faster than real time on a plain CPU. A GPU helps for larger models and for CUDA streaming mode, where Supertonic 3 reaches roughly 47 ms first-token latency.
Which languages do the new voices support?
Supertonic 3 covers 31 languages including English, Korean, Japanese, and Arabic. IndexTTS2 targets Chinese and English, Irodori-TTS focuses on Japanese, and the two MOSS-TTS models auto-detect language with an optional hint.
How is this different from ElevenLabs or other cloud TTS?
audio.cpp runs offline on your machine with no usage fees and keeps your scripts private, while cloud services bill per character and require an internet connection. Cloud tools may still lead on some high-fidelity voice-cloning tasks, so test both against your specific project.
Can I clone or convert a voice?
Yes. The engine keeps Chatterbox voice cloning and adds voice conversion in 0.3, so you can reshape a recorded take, and the zero-shot models can mimic a reference voice from a short sample.
Does it work on Windows?
Yes. The project ships prebuilt Windows binaries on its releases page, and provides build scripts for Linux and macOS with CUDA, CPU, Vulkan, and Metal backends.
