Google Transcribe

Transcribes the recording's audio with Google's speech-to-text, separating the speakers as it goes.

One of two alternative transcription steps — a journey uses this or Transcript Importer, never both. Use this one when the recording is good but the meeting provider's transcript isn't (or doesn't exist).

At a glance

Needs	The downloaded recording
Produces	The conversation's transcript, with speakers separated but unnamed
Waits when	The transcription batch is still running at Google — routinely several minutes
Re-running	Safe. But see Resetting below: re-running deletes manual speaker assignments

What it does

Extracts audio from the recording, sends it to Google Cloud Speech-to-Text for transcription with speaker separation, and writes the result as the conversation's transcript.

Speaker separation tells apart how many distinct voices there are, not who they are. Each one arrives as "Speaker 1", "Speaker 2" and so on, and someone assigns real people to them in the assign-speakers UI. That assignment matters: Content Generator refuses to write from a transcript whose speakers are unresolved.

Length limit: single-pass transcription with speaker separation caps at 60 minutes. A longer recording is rejected up front with a clear error rather than being silently truncated.

How it behaves

It submits the work and comes back for it later. Transcription takes minutes, which is far longer than a background job is allowed to run, so this step never sits and waits. Each run does two things:

1. Collects any conversation whose transcription was submitted earlier — asking Google whether that specific job finished, failed, or is still running.
2. Submits a few conversations that haven't been submitted yet (a handful per run, so one busy run can't monopolise the queue), then waits.

Over a few runs, a backlog drains steadily and several conversations transcribe concurrently. The practical effect for you: a conversation sits on this step for several minutes with no error, which is normal, and the timeline shows a "submitted" note while it's in flight.

Other behaviour worth knowing:

• It recovers from crashes. If a run dies mid-flight, the next run notices the abandoned claim, clears it, and resumes from wherever it got to — no manual intervention.
• An empty result never wipes a good transcript. If transcription comes back with no words at all, the step errors and leaves any existing transcript untouched.
• Routine waiting is silent — no notes are written while polling, so the timeline doesn't fill up with noise.

Resetting this step

A finished transcript is not silently overwritten: a re-run producing different text is refused. Resetting is the sanctioned way to replace it, and it is destructive:

• The transcript is deleted, including any manual speaker assignment. This one is not recoverable by re-running: speaker separation numbers voices arbitrarily each time, so "Speaker 2" next run needn't be the person it was last run. Keeping the old mapping would attach the wrong name to a voice, so it's discarded and speakers must be re-assigned.
• Anything derived from the transcript is deleted with it: the conversation's memory build, its extracted claims and evidence, and any answers that cited that evidence. Participants' consent decisions are not touched.
• Steps downstream that used the transcript are re-armed so they regenerate.

Nothing else is affected — the recording, the uploaded video, and publishing records stay.

Troubleshooting

Symptom	Likely cause	What to do
No movement for several minutes, a "submitted" note in the timeline	Transcription is running at Google	Normal. It collects the result on a later run.
No movement and no "submitted" note	Waiting its turn — only a few submissions happen per run	Normal on a busy backlog. Give it a few runs.
Error: recording too long	Over the 60-minute speaker-separation limit	Use the provider transcript instead, or split the recording
Error: no words / no transcript output	Silent, corrupt, or extremely short audio	Check the recording actually has audio; any existing transcript was left alone
Error mentioning credentials or a storage bucket	The Google service account or staging bucket isn't configured correctly	An administrator fixes the agent configuration, then re-run
Speakers are all "Speaker 1/2/3"	Expected — this step separates voices but can't name them	Assign real people in the assign-speakers UI, then regenerate the draft
Re-run refused	A finished transcript already exists	Deliberate. Reset the step if you really mean to replace it — you'll lose speaker assignments

Technical reference

Step type	google_transcribe
Runs after	realtimekit_downloader
Alternative to	transcript_importer (both replace the whole transcript — use exactly one)
Feeds	content_generator
Reads	infos["audio"] or config["iris.downloads"]["recording"]; its own in-flight state at config["iris.downloads"]["chirp_stt"]; step settings language, min_speakers, max_speakers
Writes	TranscriptSegment rows; infos["audio"]; config["enter-coherence"][<person_id>] speaker labels; config["iris.downloads"]["chirp_transcript"]
Needs on the agent	google.credentials (Speech-to-Text + Storage Object Admin), google.stt_bucket (US multi-region)
Needs on the server	ffmpeg, ffprobe; Speech-to-Text API enabled
Model	Chirp 3, single-pass diarization (60-minute ceiling)