Open source Swift library for on-device speech AI — ASR that beats Whisper Large v3, full-duplex speech-to-speech, native async/await

The article presents speech-swift, an open-source Swift library enabling on-device speech recognition models that outperform OpenAI's Whisper Large v3 model on the LibriSpeech benchmark. Two architectures — Qwen3-ASR, an audio language model using a large language model decoder, and Parakeet TDT, a non-autoregressive transducer — run fully on Apple Silicon without network dependency, offering faster than real-time transcription with better accuracy. These models leverage Apple Silicon's unified memory, 4-bit quantization, and hardware acceleration to achieve up to 40 times real-time speed while addressing Whisper's failure modes like hallucinations and context loss. This development signifies a shift toward efficient, privacy-preserving, high-performance speech recognition on consumer devices.