Transcription

Transcription lets your agent turn recordings into text. Attach a call, a meeting, a lecture, a podcast or a voice note, and the agent writes down what was said — with timestamps and, when the recording has several voices, a label for each speaker. It then works with the text like any other document: summarizes it, pulls out decisions, finds quotes, or reviews how the conversation went.

Short recordings come back right in the chat. Long ones take a few minutes, so the agent tells you the transcript is being prepared and replies on its own when it is ready — you don't have to wait or keep asking. In both cases the full transcript is saved as three files (.txt, .json and .srt); in a chat, the text and JSON files open on the canvas.

What you can do with it

• Attach a recording and ask for a transcript. Audio and video both work — for a video, the agent transcribes the sound. Use the attach button in the chat, the same way you attach any other file.
• Ask for more than the plain text. "Review this sales call and list the objections", "give me the three decisions from this meeting with timestamps", "quote what the customer said about pricing". The agent reads the transcript and answers from it.
• Get speaker labels. When the recording has several voices and the model can tell them apart, each line is marked Speaker A, Speaker B and so on, and the platform aims to keep the same person under the same label from the start of the recording to the end. Ask the agent to skip speaker labels if you don't want them. Labels are not names — tell the agent who is who and it can use the names from then on.
• Let the language be detected, or name it. The language is recognized automatically. If the recording is hard to recognize — noisy, or several languages mixed — tell the agent the language and it will use that.
• Download the result. Every transcript is saved as three files that share a short name the agent gives it plus a short code, for example team-sync-3f2a1c:
  • a text file (.txt) with timestamps and, when the model can tell voices apart, speaker labels, for reading and sharing;
  • a JSON file (.json) with the start time, end time and speaker of every segment, for use in other tools;
  • a subtitles file (.srt) that works as captions for the original video.
• Fill in a missing range. If part of a long recording could not be transcribed, the transcript is still delivered — the gap is marked with its time range — and you can ask the agent to fill it in. See Long recordings and gaps.
• Transcribe a recording handed over by another agent. When one agent delegates a task to another and attaches the recording, the receiving agent transcribes that recording directly — nobody has to upload it again. Only an uploaded recording can be transcribed this way; a recording an agent generated cannot.
• Use it from an interface the Builder made. If your agent has its own interface, ask the Builder to add a transcription screen: upload a recording, watch the progress, see any gaps, retry, and download the files.

The transcript files belong to the person who asked for them, like other files the agent makes for you.

Supported files and limits

Audio formats	MP3, MPGA, M4A, WAV, AAC, OGG, OGA, Opus, FLAC
Video formats	MP4, MOV, MKV, WebM, AVI, MPEG (the sound track is transcribed)
File size	The same upload limit as any other file (200 MB by default)
Length	Up to 2 hours of audio per recording

A file over the size limit or in another format is refused as soon as you send it. A recording longer than 2 hours, or a video with no sound track, is stopped when the platform starts on it, with an explanation. Nothing is charged in any of these cases. For a longer recording, split it into parts and transcribe each one.

Behind the scenes the platform converts the sound to a compact mono format and works through the recording in parts of up to about 20 minutes, a few at a time. The first part goes on its own, so that its voices can be used to keep the speaker labels the same in the later parts. That is why a long recording is faster than its running time, and why one failed part doesn't lose the others.

Long recordings and gaps

In a chat, the agent waits a short while for the result. If the transcript isn't ready by then, the agent tells you it is still being prepared and stops — it does not start the job again. When the transcript is ready, the agent continues the conversation by itself with the result.

Outside a chat — an agent run started through the API, a workflow, or a voice or messenger session — there is no conversation to continue, so a long transcription is not delivered automatically. The agent notes the job and, later in the same run, checks the job again to get the result; it does not start the job a second time. Nothing delivers the result into such a run on its own, so the agent has to check back before the run ends if it needs the transcript.

A job that waits in the queue for over an hour is cancelled without any charge. A job that runs for more than about half an hour is stopped, and you get what was finished as a partial transcript. If some parts could not be transcribed — a temporary provider problem, or your balance ran out part-way — you get a partial transcript:

• the finished parts are in the transcript as usual;
• each missing range is marked in the text, for example not transcribed, with its start and end time;
• the minutes already transcribed are already charged and are not charged again.

Ask the agent to fill the gap — "transcribe the part that's missing" — and it continues the same job with the same model, language and speaker settings, working only on the missing ranges. If the gap was caused by your balance, top up first.

How models are chosen

You don't need to know model names. The platform administrator enables the transcription models in Settings → Media models, marks one of them as the default, and describes what each is good at. The agent uses the default model unless you ask for something specific:

• Ask for speaker labels and the agent can pick a model that tells voices apart: it sees what each enabled model can do.
• Ask for a specific model by name and the agent uses exactly that one. If it isn't available, you get an explanation instead of a silent switch.
• If no transcription model has been set up yet, the agent tells you so, and your administrator needs to configure one. Nothing is charged.

Models that cannot label speakers still produce a full transcript; it just has no speaker labels. A model that cannot time individual sentences gives one timestamp per part of the recording instead.

How you turn it on

1. In the agent's Tools settings in the Builder, switch on Audio Transcription. No API keys or provider setup are needed inside the agent — the platform supplies the transcription backend.
2. Attach a recording in the chat and ask for what you need.

Audio transcription is part of every plan by default. If it isn't included in yours, the agent says that transcription needs a paid plan, and nothing is charged. For the current plans, see the pricing page; how plans unlock capabilities is described in Plans & billing.

How it is billed

Transcription is billed per minute of audio that was actually transcribed, from the same credit balance as other usage. Who pays follows the usual rules — see Choose who pays.

• Part by part. Each part of the recording is charged when it finishes, not when the job starts. Your balance is checked before each part is started, so a long recording stops when the balance runs out instead of running on; parts already in progress still finish and are charged.
• Failures are free. An attempt that fails is never charged, and a part that fails for a temporary reason is retried a few times before it is given up on. A recording that fails part-way costs only the minutes that were transcribed.
• No double payment. When you fill a gap, only the missing parts are processed and charged. Parts that already finished are never billed twice, and the new parts are charged to the same balance as the original request.

Long recordings use credits

A two-hour recording is 120 minutes of transcription. Check your balance on the Usage page before sending several long recordings.

See also Media generation for creating images and videos, and Tools & MCP for the other built-in tools.