
Local vs cloud speech recognition: choose by data flow
Compare local and cloud recognition by audio movement, network dependency, device load, and audit questions instead of treating privacy as a label.
Comparing live translation approaches — platform built-ins, desktop tools, self-hosted pipelines and interpreters — by cost and where each stops working.
The problem with most tool comparisons is that they publish a ranking without publishing the conditions.
The same setup can be excellent for Chinese to English and unusable for Chinese to Vietnamese; accurate in a quiet room and broken once background music starts. Ranking without conditions is not informative.
Every comparison states four things before it states a conclusion:
Without those four, no latency or accuracy figure is comparable to another.
Platform built-ins. Platform-side captioning and translation roll out by region and app version, with no citable specification for the LIVE side, so the boundary is whatever you see on the day. Lowest integration cost, least control: language pairs, terminology and timing are not yours to configure.
Desktop tools. The whole pipeline runs on the host’s machine. Highest control, works with virtual audio cables and OBS, but demands a capable computer and a stable connection.
Self-hosted pipelines. Wire up ASR, MT and TTS APIs yourself. Maximum flexibility, but you own the glue: segmentation, terminology, audio routing and failure recovery. That work is routinely underestimated.
Human interpreters. The highest quality ceiling and the highest cost, hard to sustain across long sessions. It fits high-ticket, low-frequency broadcasts.
This site is maintained by the team behind OBSTrans (Yanbo in Chinese), which falls into the “desktop tools” category above. Comparisons involving it say so explicitly and list the cases where it is the wrong choice.

Compare local and cloud recognition by audio movement, network dependency, device load, and audit questions instead of treating privacy as a label.

Run every candidate against one fixed recording, compute word error rate, and read the error mix to locate the real problem. Plus latency and cost basis.

Platform built-ins, desktop tools, self-hosted pipelines and live interpreters compared under stated conditions, with our own disclosure.