VoiceInput vs Superwhisper

Superwhisper is the polished Whisper desktop wrapper for English-first users. VoiceInput is built around Chinese typing speed and turns every dictation into a searchable, persona-reviewable memory — not just text dropped at the cursor.

One-line verdict

If you're an English-first user who wants the cleanest Whisper desktop wrapper, Superwhisper is excellent. If you type Chinese (or mixed Chinese-English) all day, want under-1.5-second latency, and want everything you said to be searchable next month — VoiceInput is built for that.

Side-by-side comparison

Where Superwhisper wins

• English polish. Whisper's English transcription is industry-standard. If you only speak English, Superwhisper's output is hard to beat.
• Single-purpose simplicity. No memory features, no personas — just dictate and move on. Some people prefer that.
• Mature AI prompt library. Superwhisper has built-in prompt presets for email tone, tweet style, etc. VoiceInput's AI tidy stays focused on transcription correction.

Where VoiceInput wins

• Speed on Chinese. Volcengine streaming + local typography = sub-1.5-second on Mandarin and mixed CJK-English. Superwhisper's Whisper backbone is slower for the same input.
• Memory layer. Every voice line auto-archives. A month later you can search "what did I say about onboarding?" — Superwhisper drops the text and forgets.
• AI personas. Same line re-read by Boss, Coach, Therapist, Editor. Weekly MBTI sketch from your real talk. This is a different product category, not just dictation.
• Free tier covers real usage. 100% Local and BYOK paths are free forever. Superwhisper's free tier is a 7-day trial.
• BYOK flexibility. Drop in any OpenAI-compatible endpoint (DeepSeek, Kimi, your own server). Pay model providers directly, ~$0-2/month for typical use.
• Chinese typography. Half-width spacing between CJK and Latin, brand casing (Cursor, Kimi, API stay correct), unit spacing — all handled by a local rules engine in under 5ms.

Speed: where the 1.4 seconds come from

VoiceInput's end-to-end pipeline is built for one number: time from button release to text landing at your cursor. The path:

1. Press right Option, audio streams to ASR (Volcengine) over a persistent connection — no handshake on each utterance.
2. ASR returns partial transcripts in real time. The local typography engine fixes spacing, casing, and unit formatting in <5ms (no LLM call).
3. If AI tidy is enabled, the cleaned text is sent to the LLM with a constrained prompt (three jobs only: homophones, fillers, punctuation). Confidence below 0.5 keeps the original.
4. Text injects at the cursor via Accessibility API. Clipboard fallback triggers if the target field rejects the inject.

Superwhisper's Whisper-based pipeline batches audio into chunks of 200ms-1s before processing — fundamentally a different architecture, optimized for English transcription accuracy over real-time latency.

Memory layer: the real product difference

Most dictation apps stop at "speech becomes text." VoiceInput treats each utterance as data worth keeping:

• Tool layer (SPEAK). Hold-to-talk. Text lands at the cursor. Same job Superwhisper does.
• Data layer (RECALL). Every line archives locally with the source app, timestamp, and auto-tags. Full-text search across months of dictation. Export to Markdown / JSON / CSV.
• Memory layer (REFLECT). 7 built-in personas (or your own) re-read your week. AI-picked quotes worth echoing. Weekly Big5 sketch — patterns you wouldn't see yourself.

If you mostly use voice for quick text injection, you don't need the bottom two layers. If you talk through real decisions and want them retrievable later, no Whisper wrapper has built this.

Privacy

Both apps offer a local-only mode. The difference is what leaves your Mac when you opt into cloud:

• VoiceInput cloud. Audio streams directly to Volcengine ASR — no proxy server, no audio storage. Only one number leaves: recording length in seconds (for the global pulse counter, toggleable in Settings). API keys live in macOS Keychain.
• Superwhisper cloud. Audio routes through Whisper API providers. Standard OpenAI / equivalent retention policies apply.

Who should pick which

FAQ