@kjlis Nice questions！
For discussion seek, we beef up 100+ languages via our speech reputation pipeline, together with English, French, Italian, Spanish, Jap, Chinese language, and lots of others. So long as the language is supported by means of the underlying ASR fashions, the discussion turns into searchable. Accuracy can range by means of language, audio high quality, accents, and recording stipulations, however we’ve discovered it really works rather well throughout maximum main languages.
For compound queries, sure. We don’t deal with seek as easy key phrase matching. We use semantic retrieval and reranking to know the intent in the back of a question. For one thing like:
“To find clips that include each X and Y”
clips matching each ideas would usually rank easiest, whilst clips matching handiest X or handiest Y might nonetheless seem additional down the effects if they’re semantically related. In follow, the gadget tries to optimize for the consumer’s intent reasonably than making use of strict boolean good judgment.
We’d love to listen to extra in regards to the workflows you’re enthusiastic about. That is a space we’re actively bettering.
