Fetching from the wire…
Public story · 2026-07-31 · high
The Swift menu-bar app wires an existing speech model into Core Audio's process-tap APIs to skip the cloud entirely.
Why now: Quill launched July 24 and built a following within its first week, right as on-device transcription quality has caught up to cloud-based tools.
Quill transcribes an hour of recorded audio in about 20 seconds, running entirely on a Mac's own chip, according to its GitHub repository.
That speed makes cloud transcription optional for anyone recording meetings, podcasts, or lectures. Audio never leaves the machine, so there's no upload step and nothing to trust to a remote server.
The repo went up July 24, per GitHub. It picked up roughly 3,000 stars in its first week, a sign developers want transcription that stays local.
The app is a single Swift binary that lives in the menu bar. It captures microphone and system audio as two separate tracks, using Core Audio's process taps and AVAudioEngine.
Transcription runs through Parakeet TDT 0.6B v2, a speech model FluidAudio ported to Core ML. Setup costs one thing: a roughly 600MB model download, after which everything runs offline.
It needs macOS 15 or later. A Whisper fallback is planned. The project is MIT-licensed.
None of the individual pieces here are new. Parakeet TDT and its Core ML port already existed, and so did Core Audio's process-tap APIs. Quill's contribution was wiring them into something a non-developer could run without touching a terminal. That reception measures pent-up demand for the integration, not a new model breakthrough.
Watch whether forks add features cloud transcription services charge for, like speaker diarization or search across old recordings. A free, on-device base changes what those subscriptions are actually selling.
Each link below shares sources, entities, or timing with this story.
The repo reached 9,740 stars with roughly double the next-fastest Python project's daily gain. It exposes a VAD → STT → LLM → TTS pipeline behind an OpenAI Realtime-compatible WebSocket API with every stage swappable: Parakeet TDT as default STT (Whisper, Paraformer alternativ...
deltafin, created July 28 and already at 243 stars, streams MXFP4 expert weights on demand over HTTP into a local disk cache rather than resident memory, with fused NEON kernels, Metal/MPS compute, exact reproducible decoding, and an OpenAI-compatible server. This directly con...
Blaizzy/nativ (1,163 stars, Swift, MIT, macOS 26+) comes from the mlx-vlm author and bundles that server into a SwiftUI app that discovers MLX models already in your HF cache. It exposes OpenAI-compatible chat, Responses, image, audio and model endpoints plus Anthropic Message...
+3,059 this week. Bundles Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo (350M), HumeAI TADA and Kokoro into one MIT-licensed local voice studio, with zero-shot cloning from audio samples plus 50+ preset voices across 23 languages, combining TTS...
This MIT Python plugin (+434 today) takes a video URL or local path with a question: pulls native captions first, falls back to Whisper via Groq or OpenAI, extracts frames with ffmpeg, dedupes near-identical frames to save tokens, and hands Claude timestamped images aligned to...
Alongside WASTE, gavamedia/deltafin (603 stars, created July 28) runs full K3 on a single device with an OpenAI-compatible server. Moonshot published K3's open weights July 26-27 at 2.8T parameters; within 48 hours two separate projects appeared whose entire purpose is fitting...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.