Google Cloud Speech-to-Text

Transcribe audio with Google Cloud Speech-to-Text across 125+ languages. Agents recognize short clips synchronously and submit long recordings as asynchronous jobs polled to completion, with word timestamps, speaker diarization, and phrase hints.

4 actions Custom auth available

Google Cloud Speech-to-Text website ↗ Platform Documentation ↗ Capability data verified 2026-07-25

A sales call recording lands in the team's Cloud Storage bucket from the phone system

Long Running Recognize starts an async job on the gs:// URI with speaker diarization and phrase hints for product names

Get Operation polls the job and collects the finished transcript with word timings and speaker tags

The flow scans the transcript for the pricing commitments and compliance phrases the team tracks

A sales manager reviews the flagged excerpts before anything is written to the rep's record

Confirmed findings land on the deal record in the CRM with timestamped quotes

A digest posts to the team channel linking each excerpt to its minute mark

What This Integration Enables

Speech-to-Text is where most companies' most honest data sits unread: the calls, voicemails, and meetings where customers say what they actually think. Google Cloud Speech-to-Text turns that audio into text an agent can act on across 125+ languages, and this connector is honest about the mechanics: synchronous Recognize handles roughly a minute of audio, anything longer runs as an asynchronous job against a Cloud Storage gs:// URI, and Recognize and Wait bounds its own waiting so a slow job degrades into polling instead of failing the flow. Transcription is only the front half of a voice pipeline; the back half is acting on what was said, and that is where FlowRunner routes transcripts into tickets, records, and review queues, with confidence scores deciding which path each one takes. - Transcribe voicemails and short clips inline with Recognize, straight from a URL - Run recordings up to roughly 480 minutes as background jobs with Long Running Recognize - Separate speakers in sales and support calls with diarization - Produce word-level timestamps for subtitles, chapter markers, and excerpt links - Bias recognition toward product names, SKUs, and domain terminology with phrase hints - Use the dedicated phone-call and enhanced models for 8 kHz telephony audio

Without FlowRunner

Recordings are write-only Calls get recorded, stored, and never listened to again, so the archive is a cost, not an asset

Who said what is unknowable A recording proves the call happened, not which speaker made the commitment

Jargon defeats transcription Product names, SKUs, and clinical terms come out mangled, so nobody trusts the text

With FlowRunner

Audio becomes queryable text Voicemails transcribe inline, and long recordings run as background jobs that never block the flow

Speakers are separated and timestamped Diarization tags each speaker, and word timings let a digest link a claim to its exact minute

The vocabulary is yours Phrase hints bias recognition toward your product names and terminology before the transcript is produced

Use Case Scenarios

Voicemails become tickets before the morning standup

Overnight voicemails from the support line drop into [Dropbox](/integrations/dropbox-service). For each file, the flow gets a temporary link, passes it to Recognize with the phone-call model, and logs the transcript and confidence score to [Google Sheets](/integrations/google-sheets). Transcripts mentioning outage or billing terms open a ticket in [Zendesk](/integrations/zendesk) with the audio attached, so the first human touch starts from text, not from replaying audio.

The webinar library becomes searchable

Each webinar recording uploads to a bucket via [Google Cloud Storage](/integrations/google-cloud-storage). The flow calls Recognize and Wait on the gs:// URI with word time offsets enabled, and if the job outlasts the bounded wait, the returned operation name goes back into a Get Operation polling loop instead of failing. The finished transcript publishes to [Notion](/integrations/notion) with chapter markers built from the timestamps, and a summary drafted by [Google Vertex AI](/integrations/google-vertex-ai) posts to [Slack](/integrations/slack).

Sales calls reviewed by exception

Call recordings from [Aircall](/integrations/aircall) land in Cloud Storage. Long Running Recognize runs each with diarization and phrase hints tuned to the product catalog, and Get Operation collects results on a schedule. The flow scans rep-attributed segments for discount language and contract terms, and only flagged calls reach the sales manager, each excerpt linked to its minute mark. The manager reviews minutes, not hours, and the archive starts earning its storage bill. Confirmed findings write through to the deal record in [HubSpot](/integrations/hubspot), so the commitment a customer heard on a Tuesday call is still on the record when renewal comes.

Human-in-Loop Highlight

The costly operation here is the backfill. Pointing Long Running Recognize at a call archive means a metered job per recording, billed against every minute of audio processed, and once a thousand-hour sweep starts, that spend cannot be recalled. It is also the moment feature choices get expensive: enhanced models, diarization, and multi-channel processing all change the bill across the whole archive at once. So a FlowRunner agent asked to make the archive searchable does not simply start submitting jobs. It inventories the bucket, estimates the audio hours, states which model and features the job will use, and posts the scope for approval. A person signs off on the sweep, then the agent runs it at machine speed, submitting, polling with Get Operation, and filing transcripts without further attention. The spend decision is made once, by someone accountable for it, instead of accumulating silently job by job.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

4 actions

Synchronous Transcription

Recognize Transcribes short audio synchronously and returns the transcript in the same call, from a public URL or a gs:// URI. Handles roughly 60 seconds of audio with a 10 MB cap on inline requests, and supports automatic punctuation, word-level timestamps, diarization, phrase hints, enhanced models, and multi-channel audio. The right tool for voicemails and clips, not recordings.

Asynchronous Jobs

Long Running Recognize Starts an asynchronous transcription job for long audio and immediately returns an operation name to poll. Handles audio up to roughly 480 minutes, supplied as a Cloud Storage gs:// URI, with the same recognition features as Recognize. The flow is never blocked while Google does the work.
Get Operation Retrieves the state of a long-running job: progress percentage while running, and the finished transcript with confidence, word timings, and speaker tags once done. The polling step behind every batch pipeline on this page.
Recognize and Wait Submits a long-running job and polls it to completion in a single step, for flows that should block until the transcript is ready. Waiting is bounded to fit the platform execution timeout, and a job that outlasts the cap returns timedOut with the operation name so the flow continues polling instead of failing.

Frequently Asked Questions

What can FlowRunner do with Google Cloud Speech-to-Text?

FlowRunner agents can run Recognize, Long Running Recognize, and Get Operation in Google Cloud Speech-to-Text, plus 1 more action.

Does connecting Google Cloud Speech-to-Text to FlowRunner require OAuth?

Google Cloud Speech-to-Text uses a custom authentication method to connect to FlowRunner.

Can Google Cloud Speech-to-Text trigger a FlowRunner workflow automatically?

Google Cloud Speech-to-Text doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with Google Cloud Speech-to-Text

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us