Transcription

Transcription is not only for Kilevel's own calls. A transcription job takes audio from anywhere — a call centre's archive, recorded meetings, voicemails — transcribes it, and runs structured outputs over the result. What comes out is the same shape as a call record, so everything that already reads calls reads these too.

A job is configured once and holds every setting: where the audio comes from, the language, how speakers are separated, what the model should extract, what is attached from your own systems, and how much of it may run at a time.

Sources

The four folder-shaped sources — S3, FTP/SFTP, Drive and a watched API job — can keep running rather than finishing: see Sources and watching for what each needs and how a watch is kept cheap on a folder that only grows.

Language

Fix the language, detect it per file, or transcribe multilingual audio in which speakers switch languages mid-conversation. Fixed is the most accurate when every recording is in the same known language; detection commits to one language per recording, which is what multilingual audio breaks.

Speakers

• Automatic — splits speakers by channel when the file has one channel per speaker, and by voice otherwise.
• Always diarize — separate speakers by voice, the only option for a mono file.
• Split by channel — one channel per speaker; a mono file fails rather than being guessed.
• No separation — a single stream of text.

Expected speakers is not sent to the provider. It decides whether channel splitting is possible at all, and caps the reported speaker count so a noisy two-party call is not labelled as seven people.

Diarization only ever says "Speaker 1" and "Speaker 2" — neither carries identity. Which of them is your agent is worked out separately, and every agent-specific number depends on it: see Analysis.

Analysis

Every new job starts with the Conversation Quality Core metric pack and the system output already attached, so a job created without touching the analysis section still produces comparable numbers rather than transcripts alone. Both are ordinary entries in the list and can be removed before the job starts.

There are twelve metric packs, an "explain every judgement" setting that makes the model write down its reasoning, and five metrics measured straight from the audio with no model involved. Analysis and metric packs covers all of it.

Details from your own systems

The audio cannot tell you who the agent was, which customer called, or when the conversation started — so Kilevel never guesses. Point a job at where you already keep that (the file name, a spreadsheet, or the call that pushes the audio) and every transcription carries it, as filterable columns kept separate from the analysis. See Details and linking.

Other job settings

Where the results appear

Transcriptions are conversations, so they show up beside calls and chats rather than only on their own page.

• Reports — pick Transcript in the session-type selector. The details become ordinary columns you can sort, group and filter by (agent, customer, project, stage, conversation start, tags), your own custom fields are available under Data Columns → Details, and the Job filter lists your transcription jobs. Clicking a row opens the transcript.
• Performance — every agent's conversations are counted across calls, chats and imported recordings together, with the split kept visible. Agents named only inside an imported archive get a row too.
• Collections — pick Transcriptions as a target to run a collection analysis over imported recordings, filtered by the details they were enriched with. Only recordings that finished transcribing are reachable, since the rest have nothing to analyse yet.

Results

Each file's transcript and extracted fields land in the same shape as a call record. Transcripts can be viewed two-sided or stacked, and exported individually or in bulk; everything is also readable over the API.