AI BriefingC++-based Unified Audio Inference Framework audio.cpp Released
Key point
audio.cpp, a ggml-based C++ audio inference framework, has been released, offering up to 5x faster speed compared to Python.
Details
audio.cpp, a native C++ inference framework built on ggml, has been released. This framework aims to unify audio models—each with their own individual Python environments and dependency trees—into a single runtime, session handling, CLI, and server environment.
Currently, 12 model families are officially released and ready for immediate use, with key included models as follows:
- TTS / Voice Cloning: Qwen3-TTS, PocketTTS, MioTTS, OmniVoice, Chatterbox, etc.
- ASR / VAD: Qwen3-ASR, Qwen3 Forced Aligner, Silero VAD
- Voice Conversion / Codec: Vevo2, Seed-VC, MioCodec
In particular, Vevo2 supports TTS, song generation, song conversion, and editing functions all together.
In terms of performance, on a CUDA environment (Ubuntu), it shows significant speed improvements over the Python reference implementation:
- PocketTTS: 3.68x faster in 1-shot execution
- Qwen3-TTS: 1.83x faster in 1-shot execution, 3.06x faster in Long-form execution
- Vevo2: 5.03x faster in 1-shot execution
- MioTTS: 2.73x faster in 1-shot execution
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Jevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs Detection
Rootless Linux Sandbox 'Drop' Released for Isolating AI Agents and Third-Party Programs