Superwhisper — Local Whisper Dictation with LLM Post-Processing and Per-App Profiles
Summary¶
Superwhisper (superwhisper.com, commercial) is the closest direct competitor to YazSes. [EVIDENCE src-013] It provides local Whisper-based dictation with LLM post-processing, clipboard injection, custom tone/style templates, and per-app profiles on macOS, Windows, and iOS. [EVIDENCE src-013] Key differentiators versus YazSes include: real-time partial transcription displayed in a floating overlay, LLM cleanup of disfluencies before injection, and custom per-app profiles that automatically switch behaviour depending on the active application (e.g. "code comment" style in VS Code, "casual" in Slack). [EVIDENCE src-013] Superwhisper's commercial success and developer community demonstrate product-market fit for developer dictation with LLM cleanup.
Key Findings¶
- Partial transcription overlay: Superwhisper displays in-progress partial transcription in a floating overlay before committing text to the application. [EVIDENCE src-013] This gives users real-time feedback that they are being heard and allows early abort. This feature is absent from YazSes's current implementation.
- LLM disfluency cleanup: Before injection, Superwhisper passes the Whisper transcript through an LLM to remove filler words, correct technical terminology, and improve coherence. [EVIDENCE src-013] This directly addresses the disfluency omission/insertion problems documented in src-012.
- Custom tone/style templates: Predefined templates ("formal", "code comment", "casual") shape the LLM post-processing prompt. [EVIDENCE src-013] Users can write custom system prompts for bespoke behaviours.
- Per-app profiles: Superwhisper detects the active application and automatically applies the appropriate profile — no manual switching required. [EVIDENCE src-013] This is the feature most frequently cited in positive user reviews. [HYPOTHESIS based on public community discussions]
- Supported developer tools: Explicitly supports Cursor, Claude Code, VS Code, and Aider. [EVIDENCE src-013] This signals that developer tooling integration is a deliberate product focus, not incidental.
- Platform: macOS, Windows, iOS. [EVIDENCE src-013] No Linux support — a direct competitive opening for YazSes.
- Product-market fit validation: Superwhisper's commercial existence and community engagement validate the market for developer-focused local dictation with LLM enhancement. [EVIDENCE src-013]
Relevance to YazSes Innovation Domains¶
| Domain | Relevance | Notes |
|---|---|---|
| Streaming ASR | HIGH | Partial overlay display is a streaming feature YazSes lacks; closing this gap is a top-priority UX improvement |
| SSH/remote | LOW | No SSH/remote workflow documented in Superwhisper |
| Code awareness | HIGH | Per-app profiles and code comment templates directly address code-aware dictation; YazSes needs equivalent |
| Accessibility | MEDIUM | LLM disfluency cleanup helps all users; no specific atypical-speech features documented |
| AR/VR/XR | LOW | No XR integration |
| Gaming | LOW | No gaming features |
Gaps Surfaced¶
- gap-competitor-1: YazSes lacks a partial transcription overlay. [EVIDENCE src-013] This is Superwhisper's most visible UX differentiator and should be the first streaming UX feature to implement.
- gap-competitor-2: YazSes has no LLM post-processing layer. [EVIDENCE src-013] The disfluency cleanup and tone/style templating features are unimplemented. The architecture must be designed to optionally route through a local LLM (Ollama) without requiring cloud access.
- gap-competitor-3: YazSes has no per-app profile system. [EVIDENCE src-013] This requires active window detection (xdotool windowgetactivewindow on Linux, accessibility APIs on macOS) and a profile switching mechanism.
- gap-competitor-4: Superwhisper has no Linux support. [EVIDENCE src-013] This is YazSes's primary platform advantage and should be maintained and communicated clearly.
- gap-competitor-5: The LLM cleanup latency impact on end-to-end injection time is not documented in Superwhisper's public materials — users may experience a perceptible pause between speech end and text appearance. [TODO: find source]