audio.cpp adds streaming support and Nemotron 3.5 ASR

The audio.cpp project has updated its native C++/GGML framework to include initial streaming support for speech recognition and synthesis, alongside four new ASR/STT models: Nemotron 3.5 ASR, Higgs Audio STT, VibeVoice ASR, and Hviske ASR.

• Nemotron ASR transcribed a 327.6s audio file in 2.17s offline on an RTX 5090 with 3.18% WER, while its streaming SSE path achieved the same WER with 307ms TTFT and significantly lower peak VRAM.
• VibeVoice ASR recorded the lowest WER (0.66%) but required substantially more VRAM (25833M) compared to Nemotron.
• Streaming capabilities were implemented for Nemotron ASR (via SSE) and VoxCPM2 TTS, with Higgs Audio STT streaming marked as experimental.
• The update aims to provide a unified local audio framework that avoids the need for separate Python environments for each model.

The author highlights the strong speed-to-VRAM tradeoff of Nemotron ASR as particularly suitable for local ASR service usage and invites contributions to improve streaming APIs and chunk scheduling.