Multilingual Speech-to-Text

Gladia's API converts spoken audio into text across a range of languages. A Beam agent sends a recording named in the approved rule, such as a new support call, for transcription, then checks the resulting text against a keyword the customer cares about, for example a mention of cancelling or a competitor's name, and notifies the relevant team. It relies entirely on the transcript Gladia returns, it does not interpret tone or emotion beyond the words themselves. Anything involving a transcript with low confidence, an unclear accent, or an unsupported language is routed to a human to check the audio directly.

Speaker Diarization

Gladia can separate a recording's transcript by speaker, marking who said what across a call or meeting. A Beam agent uses this breakdown to apply the approved rule, for example flagging a specific line from the customer rather than the agent on a support call, and adds the relevant excerpt to a notes file or ticket. It only labels speakers the way Gladia's diarization already has, it does not verify identity beyond that. Anything involving a call with unclear speaker separation, or more speakers than the rule accounts for, is routed to a human to listen and sort out.

Audio Event Detection

Beyond words, Gladia's API can flag audio events like silence, overlapping speech, or a change in speaking pace within a recording. A Beam agent checks these events against the approved rule, for example noting a long silence on a support call as a possible dropped connection, and adds that flag to the call's record for the team to see. It only reports what the API detects, not a diagnosis of the cause. Anything involving a pattern the rule doesn't cover, or repeated flags on the same account, is routed to a human to listen to the recording directly.