Google Cloud Speech-to-Text pricing starts at $0.016 per minute for standard V2 recognition, while V2 dynamic batch costs $0.003 per minute when delayed processing is acceptable. The advertised rate is only the starting point: Google rounds every successful request up to the next second, bills each audio channel separately and charges for processed audio even when the response contains no transcript.

This guide calculates the real cost of meetings, podcasts, and support calls, then shows how V1 and V2 billing, retries, chunking, diarisation, channel separation, Cloud Storage, and the surrounding infrastructure affect the final invoice. Prices were checked on 15 July 2026 against the official Google Cloud Speech-to-Text pricing page. All figures below are in US dollars and exclude taxes and negotiated enterprise discounts.

Google Cloud Speech-to-Text pricing at a glance

API and recognition type	Current price	Practical use
V2 standard recognition	$0.016 per minute for the first 500,000 minutes per month	Real-time, synchronous and normal batch workloads where results are needed promptly
V2 dynamic batch	$0.003 per minute	Non-urgent recorded audio that can wait for lower-priority processing
V1 standard with data logging	First 60 minutes free each month, then $0.016 per minute	Legacy workloads that meet the data logging and governance requirements
V1 standard without data logging	First 60 minutes free each month, then $0.024 per minute	Legacy V1 workloads that do not opt into data logging
V1 medical models	First 60 minutes free each month, then $0.078 per minute	Medical dictation and medical conversation workflows

The most useful budgeting formula is:

Billable seconds = the rounded duration of every successful request multiplied by its channel count, plus any repeated or overlapping audio.

Convert that total to minutes, apply the correct API rate, then add storage, compute, logging, network, and manual quality-control costs. A spreadsheet that only multiplies source-file duration by $0.016 will understate many production workloads.

---

---

V2 standard pricing uses monthly volume bands

V2 standard monthly usage	Price per minute	Price per audio hour
0 to 500,000 minutes	$0.016	$0.96
500,000 to 1,000,000 minutes	$0.010	$0.60
1,000,000 to 2,000,000 minutes	$0.008	$0.48
Above 2,000,000 minutes	$0.004	$0.24

Most developers and smaller businesses remain in the first band, so $0.96 per mono audio hour is the sensible default estimate. The lower rates only become relevant at enormous monthly volumes. At that scale, use Google Cloud Billing exports and SKU-level reporting rather than relying on a single blended rate in an application dashboard.

Google also invites very large customers to request additional volume discounts. Contract pricing can therefore beat the published table, but it should not be used in a public cost model until a quote has been agreed.

Does Google Speech-to-Text have a free tier in 2026?

Google’s recurring Speech-to-Text free allowance applies to V1, not the published V2 rate card. V1 includes 60 free minutes per account each month for standard recognition with or without data logging. The same 60-minute allowance applies to V1 medical models.

V2 standard recognition starts at $0.016 per minute from the first minute. Google may also offer new-customer Cloud credits, but those are temporary account-level credits rather than a permanent V2 Speech-to-Text free tier.

The V1 allowance saves a maximum of $0.96 per month on the $0.016 rate. For 1,000 monthly minutes, V1 with data logging would cost $15.04 after the allowance, compared with $16 for V2 standard. That small saving is rarely a good reason to design a new system around the older API. API capabilities, model availability, regional deployment and data governance usually deserve more weight than a fixed 60-minute credit.

Dynamic batch cuts the transcription rate by more than 80%

V2 dynamic batch costs $0.003 per minute, or $0.18 per mono audio hour. It is designed for recorded audio that can be processed at a lower urgency, not for live captions or a customer-facing workflow that requires an immediate transcript.

Workload	V2 standard	V2 dynamic batch	Saving
100 audio hours	$96	$18	$78
1,000 audio hours	$960	$180	$780
10,000 audio hours	$9,600	$1,800	$7,800

The operational mistake is sending every file through the fast path because it is easier to maintain one queue. A better design separates urgent and non-urgent work. Live captions, interactive voice features and immediate call summaries use standard recognition. Archived calls, media backlogs, research recordings and overnight processing use dynamic batch.

This routing decision often saves more than negotiating a small percentage discount. It also avoids moving the whole workload to a cheaper provider when only the non-urgent portion needs a lower rate.

Real cost per meeting hour

Meeting setup	Billable minutes	V2 standard	Dynamic batch
60-minute mono recording	60	$0.96	$0.18
60-minute two-channel recording	120	$1.92	$0.36
30 one-hour mono meetings per month	1,800	$28.80	$5.40

Speaker diarisation does not appear as a separate surcharge on the published rate card for Google Speech-to-Text. A mono meeting with diarisation therefore remains a single-channel workload. Accuracy can still suffer with overlapping speakers, poor microphones and similar voices, so the cheaper billing setup may create more editing work.

Separate channels are different. If the organiser and participant are recorded on separate tracks, Google bills for both channels’ durations. Paying another $0.96 per meeting hour can be worthwhile when accurate speaker attribution feeds compliance records, sales coaching or support analytics. For an informal internal meeting, mono audio plus diarisation is often the better cost trade-off.

Real cost per podcast episode

90-minute episode	V2 standard	Dynamic batch
Mono mix	$1.44	$0.27
Two billable channels	$2.88	$0.54
Four isolated channels	$5.76	$1.08

Podcast exports deserve inspection before upload. A stereo file is not automatically worth processing as two channels. Many finished episodes contain the same programme mix on both left and right channels, so downmixing to mono can halve the Speech-to-Text charge without discarding distinct speakers.

Keep isolated channels when they genuinely improve editing or attribution. Four remote guests recorded on four tracks cost four times as much to process, but clean speaker separation may remove hours of transcript correction. The cheapest API bill is not always the cheapest finished transcript.

Real cost per 1,000 support calls

Assume 1,000 calls averaging six minutes each:

Support-call setup	Billable minutes	V2 standard	Dynamic batch
Single mixed channel	6,000	$96	$18
Agent and customer on separate channels	12,000	$192	$36

Dynamic batch suits post-call analytics that can wait. It does not suit an agent-assist feature that needs words, intent or risk signals while the conversation is happening.

The channel decision should follow the use case. Dual-channel audio doubles the API charge, yet it can make agent-versus-customer analysis far cleaner than asking diarisation to infer speakers from a mixed recording. If supervisors manually fix speaker labels later, the apparent saving from mono audio can disappear quickly.

One-second rounding hurts short clips more than long recordings

Google rounds each successful request up to the next one-second increment. The effect is negligible on a 90-minute podcast and potentially large for voice commands, voicemail fragments or Internet of Things events sent as separate requests.

Consider 100,000 clips that are each 1.2 seconds long. The source audio totals 2,000 minutes, which suggests a $32 V2 standard bill. Because each request rounds to 2 seconds, Google bills 3,333.33 minutes instead, resulting in a charge of about $53.33. Request-level rounding adds more than $21 without processing any extra source audio.

Do not concatenate unrelated user audio merely to manipulate billing. That complicates privacy controls, latency, failure recovery and result mapping. Instead, include rounding in the unit economics and avoid unnecessary fragmentation inside your own pipeline.

Long-file splitting creates overlap and repeated-audio costs

Long recordings are often split to meet processing limits, to parallelise work, or to recover more cleanly from failures. Every chunk is rounded independently. Any overlap added to preserve sentence context is also billed as fresh audio.

A 90-minute podcast divided into six chunks with 15 seconds of overlap at five boundaries adds 75 seconds of duplicate audio. That costs only $0.02 at the V2 standard mono rate, but over 1,000 episodes it comes to $20. Two-channel processing doubles it to $40. Larger overlaps, more aggressive chunking and retrying whole batches magnify the effect.

The practical rule is simple: split for a technical reason, not by habit. Store chunk boundaries, overlap duration and source hashes so the billing record can distinguish original audio from deliberate duplication.

Retries can be billed twice even when the first result was never used

Google does not charge requests that end in a server error. It successfully charges processed audio, including an empty response when no speech was transcribed. The awkward case is a client timeout, lost callback or worker crash after Google has accepted the job. Submitting the audio again can create a second successful processing event.

A 5% duplicate-processing rate on the 6,000-minute support-call example adds $4.80 at the V2 standard mono rate. With two channels, the same mistake adds $9.60. The percentages are only examples, but the mechanism is common enough to deserve explicit monitoring.

For batch jobs, persist the long-running operation identifier and poll the existing operation instead of creating a replacement job. For synchronous or streaming workflows, keep a request ledger using an audio hash, source ID and attempt number. Automatic retries should distinguish transport failures from confirmed API failures.

A recurring pattern in production discussions is that the provider’s per-minute arithmetic is usually correct, while the application quietly resends, overlaps or reprocesses audio. Cost monitoring needs to follow the audio through the workflow, not stop at the upload button.

Empty responses and silence still consume budget

If Google successfully processes a request but cannot produce a transcript, the audio is still billable. Silence, corrupt speech, music-only files, muted calls, and unsupported audio conditions can therefore result in charges without usable text.

Validate duration, codec, channel layout and signal level before submission. Voice activity detection can remove long leading and trailing silence, but it should be tuned conservatively. An aggressive silence trimmer that clips the first syllable of every sentence may save fractions of a cent and create expensive transcript errors.

Cloud Storage, compute, and network charges sit outside the speech rate

Google’s Speech-to-Text invoice is only one part of a production transcription system. Batch audio is commonly staged in Cloud Storage, results may be written back as JSON, and application code may run through Cloud Run, App Engine, Compute Engine or functions. Pub/Sub queues, logging, monitoring, BigQuery analysis, encryption keys, and data-loss prevention can each add a separate SKU.

Compressed audio storage is often modest, but poor lifecycle rules can leave duplicate uploads and transcript artefacts accumulating for years. Cross-region transfers and exporting data from Google Cloud can matter more than storage itself. Keep the audio bucket, recogniser location and downstream processing in compatible regions where governance allows.

Budget these costs in separate lines:

• Speech recognition by API version, method and channel count
• Input and output storage, including retention and duplicate objects
• Compute used to upload, convert, segment and post-process audio
• Queueing, logging, monitoring and analytics
• Network transfer between regions or out of Google Cloud
• Manual correction and quality assurance

The last line is frequently the largest. A model that saves $0.50 per audio hour but creates ten extra minutes of human cleanup is not cheaper.

Diarisation versus separate channels: choose the cheaper source of accuracy

Diarisation asks the model to infer speaker changes from one mixed channel. Channel separation gives the API distinct tracks and removes much of that guesswork. Google does not list a separate diarisation fee, but it does multiply billing by the number of channels processed.

Approach	API cost	Likely trade-off
Mono plus diarisation	One billable channel	Lowest recognition cost, but more speaker-label errors in crosstalk and overlapping speech
Separate agent and customer channels	Two billable channels	Higher API cost, cleaner attribution for call analytics and compliance
Several isolated podcast tracks	One charge for every processed track	Strong editing control, but costs rise linearly with guests

Use a sample of real audio to measure correction time. If dual-channel input saves even a few minutes of manual speaker repair per hour, doubling a sub-$1 transcription charge can be economically sensible.

When Whisper or Deepgram is cheaper operationally

Google V2 standard costs $0.016 per minute at normal volumes. OpenAI’s current transcription options include $0.003 per minute for gpt-4o-mini-transcribe and $0.006 per minute for Whisper-1 or gpt-4o-transcribe. See our OpenAI Whisper API pricing guide for the full model and usage comparison.

Deepgram’s listed Nova-3 rates are also below Google’s standard low-volume rate, particularly for real-time monolingual streaming, although diarisation and key-term add-ons can narrow the gap. Price pages do not settle the decision on their own.

Workload	Likely cost winner	Reason
Large, non-urgent archive	Google dynamic batch or OpenAI mini transcription	Both reach $0.003 per minute before surrounding infrastructure and quality-control costs
Real-time voice application	Often Deepgram	Lower listed streaming rates and a workflow built around live speech can outweigh Google’s wider Cloud integration
Google Cloud-native application	Often Google	Identity, regional controls, logging and existing infrastructure may remove integration work
Multilingual batch transcription	Test Google and OpenAI on real audio	Cleanup time can outweigh a one-cent difference in API price
Self-hosted, predictable batch volume	Potentially open-source Whisper	Owned compute can beat API pricing when utilisation is high, and engineering support already exists

Google becomes expensive when a team pays the standard rate for work that could use dynamic batch, processes unnecessary channels or accepts repeated audio. Whisper becomes expensive when file handling, chunk stitching, model limitations or slower processing create operational work. Deepgram becomes less compelling when paid add-ons are required for features included elsewhere or when the organisation already has mature Google Cloud controls.

For product quality, language coverage, streaming behaviour and implementation limits rather than billing detail, read our Google Speech-to-Text review. Keeping quality and pricing decisions separate prevents a low price from becoming the sole product verdict.

A better way to estimate the monthly invoice

Build the forecast from request data rather than content labels such as “meeting” or “call”. Track these fields for each job:

• Source duration in milliseconds
• Rounded billable seconds
• Number of processed channels
• V1 or V2 API
• Standard or dynamic batch processing
• Model and recogniser location
• Original, retry or deliberate overlap status
• Success, empty result, server error or client-side failure
• Storage and compute resources triggered by the job
• Manual correction time or rejection status

Then report three figures:

1. API cost per source audio hour reveals rounding, channels and duplicated audio.
2. Platform cost per completed transcript adds storage, compute and queueing.
3. Cost per accepted transcript adds retries, failed outputs and human correction.

The third figure is the one procurement and product teams should compare across providers. It exposes the false economy of a low rate paired with poor accuracy, brittle integrations or heavy transcript cleanup.

Google Speech-to-Text cost-control checklist

• Use V2 dynamic batch for every workload that does not need prompt results.
• Downmix stereo audio to mono when channels do not contain distinct information.
• Keep separate channels where speaker attribution saves meaningful correction time.
• Trim long leading and trailing silence without clipping speech.
• Avoid excessive chunking and record deliberate overlap as duplicate audio.
• Persist operation identifiers so that workers poll existing batch jobs rather than resubmitting them.
• Validate codec, duration and signal level before sending billable audio.
• Apply lifecycle rules to source audio, temporary chunks and transcript outputs.
• Keep Cloud resources in compatible regions to reduce transfer and governance complications.
• Set budget alerts by Speech-to-Text SKU and investigate cost per accepted transcript, not only total spend.

Frequently asked questions

How much does Google Cloud Speech-to-Text cost per minute?

V2 standard recognition starts at $0.016 per minute for the first 500,000 minutes per month. V2 dynamic batch costs $0.003 per minute. Higher V2 standard volume bands reduce the published rate to $0.010, $0.008 and eventually $0.004 per minute.

Is Google Speech-to-Text free?

V1 includes 60 free minutes per account each month before paid usage begins. The published V2 pricing table does not include the same recurring free allowance. New-customer Google Cloud credits are separate promotional credits rather than a permanent V2 free tier.

Does Google charge for silence or an empty transcript?

Yes, if the audio was successfully processed. Google states that an empty response is billable because the service processed the audio even though it could not transcribe it. Requests that result in a server error are not charged.

Does speaker diarisation cost extra?

Google’s published Speech-to-Text pricing does not list a separate diarisation fee. Multi-channel audio is billed differently: each channel’s duration is added to the billable total, so a two-channel recording normally costs twice as much as a mono file of the same length.

Is dynamic batch always the cheapest choice?

It is Google’s cheapest published V2 rate, but it only fits batch audio that can tolerate lower-priority processing. Live captions, agent assistance and interactive voice products still need a faster recognition path.

What is the most common Google Speech-to-Text billing mistake?

Teams often calculate cost based on source duration while ignoring channel multiplication, request rounding, chunk overlap, and repeated jobs. Logging the billable shape of each request is more reliable than estimating from the number of uploaded files.

The practical pricing verdict

Google Cloud Speech-to-Text is reasonably priced for a production API, but only when the processing method matches the workload. Standard V2 costs $0.96 per mono hour, while dynamic batch reduces that rate to $0.18 per hour. That gap is large enough to justify separate urgent and non-urgent queues.

Channel count is the next major decision. Mono plus diarisation keeps the API charge low. Isolated channels double or multiply it, but can still reduce total cost by improving speaker attribution. Measure transcript correction time before forcing every recording into the cheapest format.

For a trustworthy forecast, track rounded seconds, channels, retries, overlap and empty responses at request level. Add Cloud infrastructure and human cleanup, then compare providers using cost per accepted transcript. That figure tells you far more than the headline per-minute price.