Presspeech vs Superscribe Lite
Both offer no-subscription, local dictation for Mac and Windows. Superscribe Lite emphasizes streaming words into the target field while you speak and offers optional cleanup through a provider you configure. Presspeech transcribes after recording, checks the original paste destination, and keeps a manual recovery path when it cannot verify that destination. These are workflow tradeoffs, not measured claims about accuracy or reliability.
| Question | Presspeech | Superscribe Lite |
|---|---|---|
| Dictation workflow | Hold a hotkey or use toggle mode, speak, then transcribe after stopping. It normally pastes only after confirming the same window is still focused; otherwise it leaves the complete transcript for manual paste. | Hold a shortcut or tap it for hands-free recording; its September 2026 launch note describes words streaming into the current field while speaking. Mac defaults to Option+Space and Windows to F9. |
| Recognition | Local recognition on each platform: multilingual Parakeet TDT v3 on macOS; Windows defaults to multilingual Parakeet when usable NVIDIA CUDA is available, otherwise English-only Whisper base.en. Other Windows models are selectable. | Choose a supported local speech model for the language you speak. The vendor's September 4, 2026 launch note for version 0.3.1 listed ten on-device language options at launch; check its current download and setup details for the latest selection. |
| Text handling | Deterministic dictionary corrections, shortcuts, filler removal, and opt-in spoken formatting. There is no generative rewrite or provider-cleanup path. | Local recognition is the default. Optional cleanup can remove filler words, change punctuation, or rewrite text using an API key the user supplies to a provider; the cleanup request sends text to that provider. |
| Privacy and network | No account or cloud transcription path. Model downloads and optional release checks use the network, and bundled-library behavior can differ by release. Review the version-specific network inventory and platform-specific macOS and Windows notices before choosing a build. | The vendor says Lite has no Superscribe account connection and keeps audio and transcripts on-device when optional cleanup is not used. Model downloads require an internet connection; optional cleanup sends text to the provider configured by the user. |
| Platforms and release status | At the 23 September 2026 check, the Mac release was 0.3.8 (signed and notarised for Apple Silicon); the Windows build was 0.1.12, an unsigned x64 prerelease. | The vendor offers separate Mac and Windows downloads. Its September 4, 2026 launch post identifies version 0.3.1; the inspected public materials do not establish a directly comparable signing or notarization status. |
| Published performance evidence | Publishes model-only p50 latency on four synthetic Mac clips; microphone capture, transcription startup, and paste are outside that measurement. | No directly comparable latency or accuracy benchmark was identified in the reviewed public materials. This does not establish a speed or accuracy winner. |
| Price | Free and MIT licensed. | Free; no account or subscription is required for Lite, according to the vendor. |
Choose Superscribe Lite if you want to see text arrive during speech, select a local model for your language, or optionally clean up text with a provider you configure. Choose Presspeech if you prefer a fixed, after-recording workflow, deterministic text handling, and explicit same-window checks with clipboard recovery when automatic paste is unsafe. Test either app in the target fields and languages you depend on; these public materials do not provide a controlled comparison of recognition or delivery quality.
Sources: Presspeech's source and platform documentation, version-specific privacy inventory, benchmark methodology, and target-app protocol; Superscribe's September 4, 2026 Lite launch note and official site (checked 23 September 2026). Superscribe feature and privacy statements above are the vendor's claims, not independent verification. No cross-product latency or accuracy winner is inferred.