# Whisper Speech-to-Text Deployment Calculator: Self-Hosted vs Cloud Cost

> Source: https://netray.co/tools/whisper-speech-to-text-deployment-calculator
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Free Tools

## Short answer

Convert monthly audio volume and GPU class into the real-time throughput and monthly cost of self-hosting Whisper, compared against a cloud speech-to-text API.

## Category

on-prem-ai

## Tool

**Kind:** calculator

**Inputs:**

#### Audio hours to transcribe per month

**Key:** audio_hours_per_month

**Type:** number

**Default:** 2000

**Min:** 1

**Max:** 500000

**Step:** 10

**Unit:** hours

**Help:** Total recorded audio duration, not wall-clock processing time. Include calls, meetings, and service recordings.

#### Deployment class and real-time factor

**Key:** realtime_factor

**Type:** select

**Default:** 12

**Unit:** x realtime

**Help:** How many hours of audio one GPU processes per hour of compute. Higher is faster and cheaper per hour of audio.

**Options:**

##### CPU only, Whisper large-v3 (0.3x realtime)

**Value:** 0.3

##### RTX 4090, Whisper large-v3 (8x realtime)

**Value:** 8

##### A100 40GB, Whisper large-v3 (12x realtime)

**Value:** 12

##### H100, Whisper large-v3-turbo (40x realtime)

**Value:** 40

#### Self-hosted compute cost

**Key:** gpu_cost_per_hour

**Type:** number

**Default:** 2.5

**Min:** 0.1

**Max:** 20

**Step:** 0.1

**Unit:** $/hr

**Help:** Blended cloud GPU rental rate, or amortized on-prem GPU cost per hour including power.

#### Cloud speech-to-text API price

**Key:** cloud_stt_price_per_min

**Type:** number

**Default:** 0.006

**Min:** 0.0005

**Max:** 0.1

**Step:** 0.0005

**Unit:** $/min

**Help:** Typical managed transcription API pricing runs $0.004 to $0.015 per minute of audio.

**Outputs:**

#### GPU hours needed per month

**Key:** gpu_hours_needed

**Expr:** audio_hours_per_month / max(realtime_factor, 0.01)

**Format:** number

**Description:** Compute hours required to transcribe your monthly audio volume at the selected real-time factor.

#### Self-hosted monthly cost

**Key:** self_hosted_monthly_cost

**Expr:** gpu_hours_needed * gpu_cost_per_hour

**Format:** currency

**Description:** GPU compute cost to run Whisper yourself at your selected deployment class.

#### Cloud API monthly cost

**Key:** cloud_monthly_cost

**Expr:** audio_hours_per_month * 60 * cloud_stt_price_per_min

**Format:** currency

**Description:** What the same volume would cost through a managed cloud speech-to-text API.

#### Monthly savings from self-hosting

**Key:** monthly_savings

**Expr:** cloud_monthly_cost - self_hosted_monthly_cost

**Format:** currency

**Description:** Positive means self-hosting is cheaper at this volume and GPU class; negative means the cloud API wins.

**Primary:** true

#### Annual savings from self-hosting

**Key:** annual_savings

**Expr:** monthly_savings * 12

**Format:** currency

**Description:** Twelve months of the monthly savings figure, before accounting for any one-time infrastructure setup cost.

**Disclaimer:** Estimates only. Real throughput varies with audio quality, batch size, and serving stack (faster-whisper, whisper.cpp, or vanilla Whisper). Cloud STT pricing varies by vendor and volume tier; check current rate cards.

## Content

**Intro:** This free Whisper speech-to-text deployment calculator sizes the GPU throughput and monthly cost of self-hosting OpenAI's Whisper model, and compares it against a managed cloud speech-to-text API. Enter monthly audio hours, your deployment class and its real-time processing factor, self-hosted compute cost, and cloud API pricing, and the tool returns GPU hours needed, self-hosted cost, cloud cost, and the monthly and annual savings delta. Whisper large-v3 and its faster turbo variant are strong enough to replace commercial transcription APIs for most enterprise use cases, and the cost crossover happens earlier than most teams expect once volume climbs past a few hundred hours a month.

**Sections:**

#### How the real-time factor drives the economics

**Body:** The real-time factor is how many hours of audio a GPU class transcribes per hour of compute time. CPU-only inference on Whisper large-v3 runs below real-time, meaning an hour of audio can take three or more hours to process, which is fine for small archival batches but useless for anything approaching production volume. A single RTX 4090 clears roughly 8x real-time, an A100 reaches about 12x, and Whisper's turbo variant on an H100 can exceed 40x real-time because it uses a distilled decoder with far fewer layers. At 2,000 hours of monthly audio, that difference is the gap between needing 250 GPU-hours and needing under 50.

**Bullets:**

- Whisper large-v3-turbo trades a small accuracy loss for a large speed gain, often the right default for enterprise volume
- Batching multiple audio files per GPU call improves throughput significantly over single-file sequential processing
- faster-whisper and whisper.cpp implementations both beat vanilla PyTorch Whisper meaningfully on the same hardware
- Noisy audio (shop floor, phone calls, overlapping speech) does not change the real-time factor but does affect accuracy

#### When self-hosting actually beats a cloud API

**Body:** Cloud speech-to-text APIs typically price between $0.004 and $0.015 per minute, which at low volume is trivial and not worth the operational overhead of running your own GPU infrastructure. The math flips as volume grows: at 2,000 hours a month, a cloud API at $0.006 per minute costs about $720, while a single A100 running at 12x real-time needs roughly 167 GPU-hours, costing well under half that on typical self-hosted or rented GPU pricing. Above a few thousand hours a month, self-hosting almost always wins on pure cost, and it wins on data control immediately regardless of volume for any audio that cannot leave your network.

#### What self-hosting adds beyond the GPU bill

**Body:** The GPU cost is the visible line item, but self-hosting also requires an audio ingestion pipeline, a queueing system for batch and streaming workloads, storage for both raw audio and transcripts, and monitoring for transcription quality drift. None of that is a large engineering lift compared to a typical RAG deployment, but it is real work that a cloud API abstracts away entirely. Budget a few weeks of platform engineering time for the initial build, and treat the ongoing operational cost as small but non-zero when comparing total cost of ownership against the API alternative.

#### How Netray deploys on-prem speech-to-text

**Body:** Netray deploys Whisper-based transcription for manufacturers who need to process service call recordings, training videos, and safety incident audio without sending it to a third-party API. We benchmark large-v3 against the turbo variant on your actual audio conditions, size the GPU deployment for your real volume with headroom for growth, and integrate transcripts directly into your ERP and quality systems so they are searchable alongside the rest of your operational data. Engagements typically start with an accuracy and throughput benchmark on a sample of your real recordings.

**Cta:** Get a Whisper deployment sized and benchmarked against your actual audio, with a real cost comparison to your current API.

**Faq:**

#### Is Whisper large-v3-turbo accurate enough to replace large-v3 in production?

**Answer:** For most enterprise use cases, yes. The turbo variant trades a small amount of accuracy, typically a fraction of a word-error-rate point on clean audio, for roughly three to four times the throughput. The gap widens slightly on noisy or heavily accented audio, so it is worth validating both models against a sample of your specific audio conditions before standardizing on turbo for everything.

#### How much GPU capacity do I need for real-time transcription instead of batch?

**Answer:** Real-time or near-real-time transcription (live captioning, active call transcription) needs enough headroom to process each stream faster than it arrives, plus buffer for concurrent streams. Size for your peak concurrent stream count rather than average, since a GPU class that handles batch volume comfortably can still fall behind if too many live streams hit it simultaneously without a queueing and overflow strategy.

#### Does self-hosting Whisper mean giving up features like speaker diarization?

**Answer:** No, but diarization is not built into Whisper itself and needs a separate pipeline stage, typically a model like pyannote layered on top of Whisper's transcription output. Budget additional GPU time for the diarization pass, and validate accuracy separately since diarization error rates on overlapping speech are a different failure mode than transcription word-error-rate.

#### At what volume does self-hosting clearly beat a cloud speech-to-text API?

**Answer:** For most teams the crossover sits somewhere between 500 and 2,000 hours of audio per month, depending on your GPU class and cloud API rate. Below that, the operational overhead of running your own pipeline usually is not worth the savings. Above it, self-hosting typically wins on cost and always wins on data control for audio that cannot leave your network for compliance reasons.

## Search terms

Whisper deployment calculator, self-hosted speech to text cost, Whisper large-v3 GPU requirements, on-prem transcription cost, Whisper turbo deployment, cloud STT vs self-hosted, speech to text infrastructure sizing

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json