Good to know

Accuracy, honestly

How it was measured

• Clip one: about 45 seconds of three players talking over each other with game music and yelling, 156 reference words, transcribed from scratch by hand. 66.7% word accuracy. Scored as a simple match rate instead, the same clip gives the 68.4% figure this site quoted before: an illustrative internal stress test on this one hard clip, not a general accuracy rate. Same clip, different scoring, not a change.
• Clip two: 25 seconds of three-player proximity chat, 51 words, corrected in the app and marked as checked. 78.4% word accuracy.
• Who checked them: the Substrike developer, on his own recordings, with one microphone setup. No independent reviewer yet.
• Word timing: after re-timing, the typical word starts about 64 ms from where it was spoken, measured on clip one (124 words). About a quarter of words were still more than 0.25 seconds off.
• Model choice (same one hard clip, not a rate): on clip one, the recommended model scored a 68.4% match rate against 63.8% for the largest model and 52% to 66% for three cloud speech services. On a single clip, gaps that small are not settled.
• Not yet measured: clean single-speaker speech, other microphones, accents and games, the smaller speech models on game audio, and processing time or memory use across PCs. The bar before quoting a headline figure is 30 clips across 3 or more games.

What the safety net catches, and what it does not

• The AI cleanup flags words the speech model was unsure of. A clip with more than three unsure words, or one the cleanup could not check, is held for you. One to three still exports, with the count shown on its row.
• Confident mistakes and missing words are not flagged, so a clip that exports on its own has not been proven correct. Watch it through before posting anything important.
• Cleanup is an optional 1.1 GB download. Without it, a batch holds every clip for review by default rather than exporting it as if it were checked. You can switch that off with “Export even if AI cleanup can’t run”.
• On clip one, turning cleanup on did not change the match rate. Its job is flagging, not guaranteed fixes.
• Names, slang and game jargon are the most common misses. The names and jargon field helps the model spell the names you give it.
• Speakers are grouped automatically, and game audio or overlapping speech can produce wrong or extra speakers. Speaker detection was tuned on a handful of the developer’s proximity-chat clips.
• Without the 378 MB timing model, captions still work but word timing is less precise. The app offers a one-click download.

Experimental features

These work, but sit outside the core workflow and have had little validation.

• Highlights finder. Ranks moments in a long recording by how loud and excited the audio gets. It does not know what a kill, a win or a joke is.
• Death detector. Looks for phrases in the transcript. Not validated across games or audio setups.
• Caption translation. Uses the local cleanup model to translate captions into other languages. Quality not measured, so check translated captions before posting.