Speech Swift beats Whisper Large v3

This is more than a “Whisper replacement” headline: it’s evidence that architecture choice plus sane quantization can matter more than brute-force model size on-device.

• –Qwen3-ASR 1.7B 8-bit slightly outperforms Whisper Large v3 on LibriSpeech, which makes the LALM case look real rather than theoretical.
• –Parakeet TDT’s non-autoregressive design is the cleanest argument here for deterministic local speech inference with fewer hallucination risks.
• –The 4-bit multilingual collapse is the practical warning sign: if you serve non-English users, aggressive quantization can wipe out the gains.
• –Reproducibility matters a lot here; benchmark scripts and a shipping Swift library make this useful to builders, not just paper readers.
• –The project is especially relevant for Apple Silicon developers who want CoreML/MLX speech pipelines without cloud dependencies.