audio.cpp 0.5 Expands Voice AI

audio.cpp 0.5 has added a wide range of new voice AI models and platform features.

• DramaBox: A prompt-oriented voice acting model based on the LTX-2.3 audio architecture, controlling emotion, tone, laughter, sighs, pauses, transitions, and speaker behavior.
• Confucius4-TTS: Provides multilingual voice transfer functionality that synthesizes speech in other supported languages based on a reference voice.
• RVC voice conversion
• BS-RoFormer vocal separation
• Added GLM-TTS, Kroko ASR, Parakeet-TDT, Inflect Micro v2, Fun-ASR-Nano

Fun-ASR-Nano was developed by the official FunASR team, and audio.cpp has also been registered on the official FunASR distribution platform.

On the infrastructure side, HIP/ROCm support for AMD GPUs has been introduced at an early stage, and Metal performance on Apple Silicon has also been improved. Real-time PCM input and cleaned-up streaming caption deltas have been added to the server/streaming path, making real-world service integration easier.

Some model integrations and non-CUDA backends still have room for optimization, and the project has presented performance improvements, backend testing, documentation, Web UI, and production deployment feedback as tasks for community contribution.