In short
- OpenAI's transcription tier costs about $0.006 per minute — roughly £0.28 per hour of audio — while open STT models hosted per hour of audio start at £0.03/hour, a gap of nearly 10x.
- A 40-minute podcast costs about £0.02 on faster-whisper-base hosted, versus roughly £0.19 on OpenAI's per-minute tier.
- Three open families cover most needs: faster-whisper (the safe default, tiny to large-v3), Qwen3-ASR (multilingual, small) and Parakeet TDT 0.6B v3 (fast English, best run locally for now).
- The big APIs' premium buys diarisation, formatting and SLAs — worth it if you need those bundled, not if you just need accurate text.
- If you own a GPU or a Mac, all of these run locally for free; self-serve gives you the same API against your own hardware at no cost.
Search for "whisper api cost" and you'll mostly find two things: OpenAI's per-minute price list, and a wall of "just run Whisper locally" tutorials. Both are fine answers, but neither tells you what transcription actually costs when you have real volumes of audio — a weekly podcast, a lecture series, a backlog of meeting recordings.
Here's the short version. OpenAI bills transcription per minute, at around $0.006 per minute — roughly $0.36 per hour of audio. Open models — the same Whisper weights, plus newer, smaller options — are hostable for a fraction of that, and when they're billed per hour of audio the maths becomes trivially legible. Let's put actual numbers on it.
What speech-to-text actually costs: per minute vs per hour
The big commercial STT APIs — OpenAI, Google, AssemblyAI, Deepgram — all price per minute of audio. OpenAI's transcription tier sits around $0.006/minute; the others are in a similar ballpark, with some cheaper entry tiers depending on the model and volume commitment. Per-minute pricing is fine for the occasional snippet, but it gets opaque fast:
- A 40-minute podcast = $0.24 on OpenAI's tier. Fine.
- A 10-hour archive = $3.60. Now you're doing arithmetic.
- A year of weekly episodes plus meeting recordings plus voice notes — at that point you're guessing.
The conversion is simple: per-minute price × 60 = per-hour price. $0.006/min ≈ $0.36/hr ≈ £0.28/hr at typical exchange rates.
Per-hour-of-audio pricing — the model used on Inpherio's marketplace, where hosts set their own prices — makes the cost legible at a glance. You see £0.03/hour and you know a 10-hour archive costs £0.30. No calculator, no surprise. That £0.28/hr big-API rate is nearly 10x the cheapest open-model hosting.
The full billing mechanics — how hosts set prices and what the platform fee is — are on the pricing page.
The open STT landscape in late 2026: three families worth knowing
Most content about open speech-to-text stops at Whisper. That was reasonable in 2023; it's leaving good options on the table now. Three families cover almost everything:
Whisper (OpenAI, MIT licence). Still the default. The original PyTorch repo is slow, so almost nobody runs it directly — the standard routes are faster-whisper (a CTranslate2 reimplementation) and whisper.cpp (the llama.cpp-style C++ port). Sizes run from tiny (39M parameters) to large-v3 (1.5B). If you want the safe pick, it's this.
Qwen3-ASR (Alibaba). 0.6B and 1.7B variants with strong multilingual accuracy in a small package. If your audio isn't English-first, this is where to look.
Parakeet TDT 0.6B v3 (NVIDIA). English-focused, built on a token-and-duration transducer architecture designed for speed on modest hardware. For clean English speech at volume, it's a strong local option — more on running it yourself below.
The trade-off in one line: Whisper large-v3 is still the accuracy ceiling for genuinely hard audio; the small models win on speed and cost. Qwen3-ASR and the smaller Whisper sizes sit in the sweet spot for most podcast and meeting work.
If you're self-hosting, the same logic that applies to chat engines applies here — as we found comparing LM Studio and Ollama, the engine matters less than whether the model fits in memory. For STT, the two engines worth knowing are whisper.cpp and Speaches (the faster-whisper server) — both are in the connector's supported engines.
Price comparison: the big APIs vs open models per hour of audio
Here's the table the search query is really after. The OpenAI figure is its published transcription tier; the open-model figures are live host-set prices on the marketplace — check there for current numbers, since hosts can change them.
| Model / service | Per minute | Per hour of audio | Notes |
|---|---|---|---|
| OpenAI transcription tier | ~$0.006 | ~$0.36 (~£0.28) | Managed, includes uptime and support |
| Other big APIs (Google, AssemblyAI, Deepgram) | similar ballpark | varies | All per-minute; some entry tiers cheaper than OpenAI's |
| faster-whisper-tiny / base / small / medium | — | from £0.03 | Host-set; same weights as OpenAI's Whisper |
| Qwen3-ASR-0.6B | — | from £0.03 | Multilingual, 0.6B |
| Parakeet TDT 0.6B v3 (ONNX) | — | from £0.03 | Fast English |
| faster-whisper-large-v3 | — | from £0.083 | The accuracy ceiling |
| Qwen3-ASR-1.7B | — | from £0.083 | Multilingual, larger |
The gap: OpenAI's per-minute tier works out to roughly 10x the cheapest open-model hosting per hour of audio. Some other providers' entry tiers are cheaper than OpenAI's, so measure the gap against the specific provider you're using — but none of the big per-minute APIs get close to £0.03/hour.
The honest caveat: the premium buys something real. The big APIs bundle diarisation, formatting (paragraphs, punctuation tuning), uptime SLAs and support. If you need speaker labels out of the box and can't afford to build that pipeline, the extra ~£0.25/hour is a reasonable bill. If you just need accurate text — which is most podcast, meeting and note-taking work — you're paying for things you're not using.
Worked example: a 40-minute podcast for about 2p
Concrete numbers. A 40-minute episode:
- faster-whisper-base at £0.03/hour: 40 minutes = 0.667 hours × £0.03 = £0.02. Two pence.
- OpenAI's ~$0.006/min tier: 40 × $0.006 = $0.24 ≈ £0.19. About 10x.
Scale it to a year of a weekly podcast: 40 minutes × 52 weeks ≈ 34.7 hours of audio. That's £1.04 on the open route versus roughly £9.80 on the per-minute tier. Neither breaks the bank — but this is one show, and most people transcribing things have more than one source of audio.
And where accuracy genuinely matters — noisy recordings, heavy accents, multiple overlapping speakers — stepping up to large-v3 at £0.083/hour still keeps the year under £3.
The call itself is an OpenAI-compatible endpoint, so switching from another API is a base-URL change, not a rewrite:
from openai import OpenAI
client = OpenAI(
base_url="https://api.inpherio.co.uk/v1",
api_key="YOUR_KEY",
)
with open("episode.mp3", "rb") as f:
transcript = client.audio.transcriptions.create(
model="Systran/faster-whisper-base",
file=f,
)
print(transcript.text)
Quick-start details are on the developers page; full endpoint parameters are in the API reference.
Which model should you pick for your audio?
| Your audio | Recommended model | Price tier |
|---|---|---|
| Clean English podcast | faster-whisper-small | £0.03/hr |
| Meetings, one or two speakers | faster-whisper-small or medium | £0.03/hr |
| Multilingual or accented speech | Qwen3-ASR-0.6B (cost) or 1.7B (accuracy) | £0.03–0.083/hr |
| Noisy field recordings, maximum accuracy | faster-whisper-large-v3 | £0.083/hr |
| Large batch backlogs | faster-whisper-tiny or small | £0.03/hr |
On batch backlogs: the small faster-whisper models transcribe much faster than real time on modest hardware — a well-established property of the CTranslate2 implementation — so even a 10-hour archive is an afternoon of compute, not an overnight job. All eight STT models on the marketplace and their live prices are listed on the marketplace.
What we'd pick: for English speech, start with faster-whisper-small — if the output looks good on your actual audio, you're done at the cheapest tier. For anything multilingual, Qwen3-ASR-0.6B first, 1.7B if it stumbles. Reach for large-v3 only when you've seen the smaller model fail.
Or run it yourself: the free route
This is the answer most of this blog's readers already know: all three families run locally, and if you own the hardware, transcription is free at any volume.
- faster-whisper: run it via the Speaches server (OpenAI-compatible API out of the box) or whisper.cpp for a leaner C++ route.
- Qwen3-ASR and Parakeet: both run from their official repos, with ONNX builds available for local serving — Parakeet's ONNX build is a good fit if you want it behind a simple endpoint on your own box.
Hardware reality check — and the same advice we gave when comparing vLLM and Ollama: test on your own machine before believing any published throughput number.
- faster-whisper-small runs comfortably on a 6GB GPU or an 8GB Mac
- large-v3 wants roughly 10GB VRAM or a 16GB+ Mac for comfortable speeds
- Qwen3-ASR-0.6B and Parakeet 0.6B fit below that — a modest gaming GPU handles them easily
If you already have the hardware, self-hosting wins on cost at any volume. The hosted route only makes sense when you don't want to own uptime, or you need to burst beyond what your machine can chew through.
There's also a middle path: Inpherio's self-serve lets you call your own machines through the same API for free — the hosted prices above are what you'd pay only when a request routes to someone else's hardware. Same code, zero cost, your box.
And the flip side of the same coin: if your GPU sits idle most of the day, hosting an STT model is one of the easiest ways to earn from it. Audio requests are short, predictable and easy to serve alongside other work — details on becoming a provider.
Questions people ask
How much does the Whisper API cost per hour of audio?
OpenAI's transcription tier is around $0.006 per minute, which works out to roughly $0.36 — about £0.28 — per hour of audio. Open models hosted per hour of audio start at £0.03/hour, a gap of nearly 10x.
Is faster-whisper as accurate as OpenAI's Whisper API?
faster-whisper is a reimplementation of the same Whisper weights, so accuracy is essentially identical for a given size. The differences are speed and memory use, not transcription quality.
What's the cheapest way to transcribe a podcast?
A hosted open model like faster-whisper-base at £0.03 per hour of audio, or running faster-whisper yourself on a 6GB GPU or 8GB Mac for free.
Can I run Whisper locally instead of paying for an API?
Yes — faster-whisper, whisper.cpp, Qwen3-ASR and Parakeet all run locally. Small models fit a 6GB GPU or 8GB Mac; large-v3 wants roughly 10GB VRAM or a 16GB+ Mac.
Do per-hour and per-minute STT prices include diarisation?
Generally no — cheap tiers on both sides bill for transcription only. Diarisation (who spoke when) is usually a separate, pricier tier or something you add yourself.
Which STT model is best for non-English audio?
Qwen3-ASR (0.6B or 1.7B) or Whisper medium/large-v3. Whisper large-v3 remains the accuracy ceiling for difficult multilingual audio.
Models in this post — live on Inpherio
Have a GPU, a Mac or a spare PC?
Let it earn while you sleep.
Run open models for people who need them. You set the price and keep everything but the 15% fee.
Become a hostBuilding with AI?
Every open model, one API.
OpenAI-compatible. Priced by the people who run the hardware.
Start free See the marketplaceRead “What speech-to-text actually costs: per-minute APIs vs open models” at https://inpherio.co.uk/blog/what-speech-to-text-actually-costs-per-minute-apis-vs-open-models