Voz hits 5-20x faster on-device speech-to-text by heavily optimizing for Apple's ANE

pcuenq · x · 2026-09-04

• Developer finnvoorhees showed that heavily optimizing models for Apple's Neural Engine (ANE) can speed up on-device speech-to-text by 5-20x on iPhones and Macs.
• Voz is a rearchitected version of NVIDIA's Parakeet and can transcribe 30 minutes of audio in about 5 seconds, entirely locally.
• The takeaway: targeting the ANE directly is a viable path to fast, private, offline transcription on Apple silicon.

Original post →

More from Infra

• DeepSeek reportedly bought 160,000 Huawei Ascend 950DT chips for new 1GW Inner Mongolia data center — ns123abc · 2026-09-04
• Morgan Stanley: 20% of Star Market IPOs now target China's 'chokepoint' tech — pstAsiatech · 2026-09-04
• DeepSeek plans 160,000-chip Huawei Ascend 950DT cluster in Inner Mongolia, Bloomberg reports — kimmonismus · 2026-09-04
• Zeiss executive: China about 15 years behind on EUV lithography tools — pstAsiatech · 2026-09-04
• Minisforum MS-S1 MAX P495 listing hints at €7999 price, double the original MS-S1 MAX — fairydreaming · 2026-09-04
• First community MLX 4-bit benchmarks for K2-Horizon-MoVA-36B hit 49.1 tok/s locally — DerTomsn · 2026-09-04

Infra channel →