Why offline dictation feels faster - an interactive look                        Skip to main content




Voice Type


Pricing
Learn
Enterprise
Trust
Blog


Interactive
Why offline dictation feels faster
Short phrases flatter the cloud. Long sessions expose it. Experiment with your own network and see where the time goes.
← Back to Blog  Home


Use the interactive timeline below to compare a cloud dictation workflow (upload → provider processing → optional rewrite) with Voice Type’s on-device streaming. Adjust network latency and bandwidth to see how each phase contributes to total time.

Scenario   Short (≈5s speech)  Medium (≈2 min)  Long (≈30 min)
Network profile   Office fiber - RTT 20 ms, ↑40 Mbps  Café Wi‑Fi - RTT 60 ms, ↑5 Mbps  Airport + VPN - RTT 120 ms, ↑2 Mbps  My network (measured) - RTT 60 ms, ↑5 Mbps   Use my network
Cloud STT speed   Whisper Large v3 - 189×  Whisper Large v3 Turbo - 216×  Distil-Whisper Large v3 - 250× (EN)  Custom - 200×
Compressed upload (≈128 kbps)    Cloud streaming (overlap upload+ASR)    LLM rewrite (BYOK)
Replay animation

Cloud: streaming + LLM rewrite
Total: 4.4 s





LLM rewrite (proxy)

Handshakes 240 ms     Service overhead 350 ms     Streaming (upload+ASR) NaNm NaNs     LLM rewrite (proxy) 700 ms

On‑device: finalize + LLM rewrite
Total: 3.3 s


Finalize last ~30s

LLM rewrite (BYOK)

Finalize last ~30s 2.5 s     Handshake (rewrite) 120 ms     LLM rewrite (BYOK) 700 ms


Legend
Handshakes     Service overhead     Upload     Transcribe     Streaming (upload+ASR)     LLM rewrite (proxy)     Finalize last ~30s     LLM rewrite (BYOK)

Assumptions (realistic, simplified)

Streaming can overlap upload with ASR; file upload cannot.

Handshake per hop ≈ 2×RTT (DNS+TLS+warmups). Cloud path includes ASR hop + proxy hop; BYOK uses a single hop.

Cloud ASR set above (default 200× real‑time). On‑device shows only the last ~30s finalize (≈2.5 s).

Same rewrite speed for both paths (≈1200 tok/s); proxy adds only hop latency.



Short phrases (5–15 seconds): handshakes dominate cloud flows. On-device avoids them entirely.

Long sessions: upload size and proxy hops compound latency. On-device streams live and only finalises the last ~30 seconds when you stop (about 2–3 seconds on an M1 Mac).

With bring-your-own-key rewrites, Voice Type sends text directly from your Mac to your chosen provider in a single hop. Audio stays local.







Voice Type


Learn

All guides

Voice Type vs Apple Dictation

Dragon alternatives

For writers

For developers

Notion on Mac

Latency demo

Press kit


Company

Enterprise

Trust Center

Pricing

Blog

Company

Terms of service

Privacy policy

Contact us







© 2025 Careless Whisper Inc.