Yap & Qwen Scribe: Two New Open-Source macOS Dictation Tools
30 Jul 2026
Two independent open-source projects have surfaced this week that tackle the same problem from different angles: private, on-device voice dictation for macOS. Yap, built by Frigade, leans entirely on Apple's built-in speech APIs with no model download required. Qwen Scribe takes a different route, running a local Qwen3-ASR model for transcription and system-wide dictation. Neither tool sends audio to the cloud, but they diverge sharply on compatibility, maturity, and security posture.
Yap: lightweight, but locked to macOS 26
Yap is a minimal, open-source dictation tool written in roughly 3,000 lines of native Swift. It ships as a 4 MB app that idles at around 60 MB of memory — a notably small footprint for a speech tool. It's installable via Homebrew (brew install --cask frigadehq/tap/yap), triggered with the default shortcut ⌘⇧D, and released under the MIT license.
What makes Yap distinctive is that it requires no separate model download at all. Instead, it taps directly into macOS 26 (Tahoe)'s new SpeechAnalyzer and SpeechTranscriber APIs for on-device speech-to-text. According to Yap's own benchmarks, Apple's underlying speech model outperformed Whisper Small on both word error rate (2.12% vs. 3.74% on clean audio, 4.56% vs. 7.95% on noisy audio) and speed (roughly 3x faster across 5,559 LibriSpeech clips).
The catch: Yap only works on macOS 26 or later, since it depends entirely on Apple's proprietary SpeechAnalyzer APIs. That ties its functionality — and future updates — to Apple's OS release cadence, and could meaningfully limit its usable audience until broader adoption of macOS 26 takes hold. The report includes no data on how widely macOS 26 has actually been adopted, so the size of Yap's addressable market today is unclear.
Qwen Scribe: broader compatibility, earlier-stage security
Qwen Scribe, an independent community project (not affiliated with Alibaba Cloud, the Qwen team, or Apple), is positioned as a private, local transcription and system-wide dictation tool for Apple Silicon. It supports macOS 14 or newer plus Python 3.12 or newer — a much lower bar than Yap's macOS 26 requirement.
Users can choose between two Qwen3-ASR model sizes depending on their hardware: the 1.7B model for higher accuracy (requiring roughly 3.4 GB of unified memory) or the 0.6B model for speed (roughly 1.2 GB). The tool supports drag-and-drop transcription of audio and video through a local web interface running at 127.0.0.1:8990, plus system-wide dictation activated by holding the right Command key in any text field. Additional features include automatic language detection, optional forced language, vocabulary hints, word timestamps, and SRT export, along with local transcript history with delete controls. It's licensed under Apache-2.0.
Qwen Scribe is currently at version v0.1.0-beta.1, and two caveats stand out. First, saved transcripts are stored as readable, unencrypted JSON — a potential data exposure risk for anyone dictating sensitive content. Second, the project doesn't yet ship signed or notarized binaries; that's planned for v0.2, but until then, installation may raise trust or security friction for some users. The report offers no word-error-rate comparison between Qwen Scribe's Qwen3-ASR model and Apple's speech model or Whisper, so there's no way yet to judge accuracy head-to-head.
Why founders should care
- The appearance of two independent, open-source, on-device dictation tools in the same window likely signals growing developer appetite for privacy-focused, offline-first speech-to-text on macOS — a trend founders building productivity or voice-driven products may want to watch.
- Apple's native speech APIs beating Whisper Small on both accuracy and speed in Yap's benchmarks suggests founders building macOS voice features may be better served evaluating first-party APIs before defaulting to third-party models — though these figures come from Yap's own testing and haven't been independently verified.
- Qwen Scribe's beta status and unencrypted transcript storage mean founders considering it for production use should probably budget time to assess security readiness — particularly for any workflow touching sensitive dictation content — before shipping it to customers.
- Yap's macOS 26 requirement versus Qwen Scribe's macOS 14+ support is a meaningful compatibility trade-off: founders prioritizing near-term reach across the installed Mac base may find Qwen Scribe's broader OS support more practical today, even with its added security caveats.
- Both projects' permissive licensing (MIT for Yap, Apache-2.0 for Qwen Scribe) plausibly lowers the barrier for startups to fork or embed dictation functionality directly into their own products, rather than building speech infrastructure from scratch.
What's still unknown
The report notes several open questions: neither project's adoption (downloads, GitHub stars) is documented, Frigade's business rationale for releasing Yap isn't detailed, and no independent benchmarks confirm either tool's reported performance. Founders evaluating these tools for production use should treat the current figures as vendor-reported until third-party validation emerges.