Summary
On networks where the TLS handshake to the provider is slow, every dictation pays the handshake
cost twice — once for transcription and once for LLM polishing — because connections are not
kept alive between dictations. On this machine that is the single largest latency term.
Measurements (Windows 11, OpenLess 1.3.18, BYOK OpenAI)
Two requests in a row, the second reusing the connection:
request 1: tcp connect 0.125s | TLS handshake 4.194s | total 4.410s
request 2: tcp connect 0.000s | TLS handshake 0.000s | total 0.183s
The TLS handshake alone costs ~4.2 s; a reused connection answers in 0.18 s, 24x faster.
Supporting numbers:
- DNS resolution is fine: 259-369 ms, same as any other host. Not the bottleneck.
- TCP connect is fine: 0.125 s.
chatgpt.com behaves identically (~4.1 s on first hit). The ChatGPT website feels instant only
because the browser keeps one connection open for the whole session.- Upload is not the bottleneck: 300 KB takes 2.3 s warm vs 5.4 s cold. The delta is the handshake.
This looks like SNI-based traffic inspection on the ISP side. It is outside the control of the app,
but the app decides how often that cost is paid.
What it looks like in the app
Avg per segment on this machine showed 9.5 s; individual dictations 3.1-5.7 s. The channel
health check for the LLM provider reported 5292 ms. Local ASR was tried as well (Foundry Local and
sherpa-onnx, several Whisper sizes), but with a cloud provider the handshake cost is paid on every
dictation and dominates everything else.
Requests
- Keep provider connections alive between dictations — a connection pool with a multi-minute
idle timeout instead of closing after each request.
- Warm up the connection when the app starts, and after the provider or model is changed,
so the handshake is paid once per session instead of before each dictation.
- If a request timeout exists, make it configurable. With a 4 s handshake plus a long
recording, a fixed short limit produces confusing timeout errors.
Why this is worth doing
For users behind traffic inspection this term is larger than model choice or prompt size.
Two connection settings would cut several seconds off every dictation without changing any
product behaviour.
Related: #994 (plain-text insertion mode).
Summary
On networks where the TLS handshake to the provider is slow, every dictation pays the handshake
cost twice — once for transcription and once for LLM polishing — because connections are not
kept alive between dictations. On this machine that is the single largest latency term.
Measurements (Windows 11, OpenLess 1.3.18, BYOK OpenAI)
Two requests in a row, the second reusing the connection:
The TLS handshake alone costs ~4.2 s; a reused connection answers in 0.18 s, 24x faster.
Supporting numbers:
chatgpt.combehaves identically (~4.1 s on first hit). The ChatGPT website feels instant onlybecause the browser keeps one connection open for the whole session.
This looks like SNI-based traffic inspection on the ISP side. It is outside the control of the app,
but the app decides how often that cost is paid.
What it looks like in the app
Avg per segmenton this machine showed 9.5 s; individual dictations 3.1-5.7 s. The channelhealth check for the LLM provider reported 5292 ms. Local ASR was tried as well (Foundry Local and
sherpa-onnx, several Whisper sizes), but with a cloud provider the handshake cost is paid on every
dictation and dominates everything else.
Requests
idle timeout instead of closing after each request.
so the handshake is paid once per session instead of before each dictation.
recording, a fixed short limit produces confusing timeout errors.
Why this is worth doing
For users behind traffic inspection this term is larger than model choice or prompt size.
Two connection settings would cut several seconds off every dictation without changing any
product behaviour.
Related: #994 (plain-text insertion mode).