Notes on real-time speech translation, simultaneous interpretation, and running multilingual conversations.
A bot that joins your call as a participant needs to be admitted by the host and shows up for everyone; a tool that listens through one person's device audio needs neither, at the cost of only hearing what that one device can hear.
The best automatic summaries aren't a shorter version of the transcript — they answer three specific questions: what was decided, who owns what, and what happens next.
Real-time speech translation listens to speech as it's spoken and produces translated text (and often audio) within a few seconds — continuously, not after you finish a sentence and hit a button.
On a video call, simultaneous interpretation is a short pipeline running continuously in the background: audio in, transcribed speech, translated text, and captions or synthesized speech out — fast enough that it never asks the conversation to pause and wait for it.
Chinese and English put information in different places in a sentence — Chinese often saves the key verb or qualifier for later than English would — which means a simultaneous system sometimes has to wait for more of a Chinese sentence before it can commit to an English translation, or vice versa.
The words most likely to matter in a business conversation — a budget figure, a client's name, a product term — are also the words general-purpose translation is most likely to get subtly wrong, which is why they deserve deliberate handling rather than being left to a model's best guess.
"No audio stored" means speech is held in memory only for as long as it takes to transcribe and translate it, then discarded — nothing is written to disk, and there's no recording to request, leak, or subpoena later, because it never existed as a file.
The honest framing isn't "which is better" but "which conversations actually get an interpreter today" — AI translation's real competition is usually no interpretation at all, not a booked human professional.
Most of what makes a multilingual meeting hard isn't the language gap itself — it's long uninterrupted turns and unexplained jargon, both of which are fixable with meeting habits, independent of whatever translation tool is running.
The real tradeoff isn't quality — it's reach: a browser tool works instantly for anyone with a link, while an installed app can do things a browser can't, like listening to audio from other applications on the same device.
A glossary doesn't need to be exhaustive to be useful — ten to twenty terms that are specific to your company or industry, listed with their intended translation, is enough to fix the errors that actually recur in practice.
Two specific patterns cause most of the numeric and grammatical errors in Chinese-to-English translation: measure words, which don't translate to anything and must be dropped, and large numbers, which are grouped differently in the two languages and need conversion, not word-for-word translation.
Live captions and live translation solve different problems — captions make speech readable for someone who can already understand the language spoken; translation makes it understandable for someone who can't.
Latency in real-time translation isn't just a speed metric — it's the difference between a tool that feels like part of the conversation and one that feels like reading minutes from a meeting that already ended.