OpenAI-compatible
Drop-in /v1/audio/transcriptions. Change one base_url and keep your existing SDK, prompts and pipelines.
OpenAI-compatible · speech-to-text live
The elemental inference API.
Fast, affordable AI inference behind one clean API. Start with speech-to-text — drop-in compatible, from ~$0.05 per hour of audio, 230× realtime, across 25 languages. LLM, TTS and realtime next.
Backed by the Google for Startups Cloud Program
support-call.mp3transcribing
12.4s audio · 54 ms processing · 230× realtime
230×
realtime on long audio
~$0.05
per hour of audio
25
languages, auto-detected
~1s
round trip for a short clip
Why Inferentium
Inference, refined to the element.
Everything you need to ship audio intelligence — priced like a commodity, engineered like a product.
230× realtime
A 90-minute recording transcribes in about 25 seconds on our own GPU fleet — long audio amortises the fixed cost of a request. Batched and autoscaled, so throughput holds under load.
25 languages
Automatic language detection with consistent accuracy across English, Spanish, Portuguese, French, German, Japanese and more.
Prepaid credits
Pay per second of audio. No subscriptions, no minimums. Top up with a card and spend down transparently.
Sync & async
Upload and wait, or submit a job and get a signed webhook when it is done. Built for clips and for hours-long files.
One API, many models
Speech-to-text today; LLM, text-to-speech and realtime streaming on the same keys, credits and billing next.
How it works
Live in three steps.
No infrastructure to run, no models to host, no GPUs to babysit.
01
Create a key
Sign up and generate an API key in seconds. Free credits to start — no card required.
02
Change one line
Point your existing SDK at our base URL. The rest of your code keeps working exactly as it is.
03
Ship
Transcribe a clip in about a second, or hours of audio in seconds. Pay only for what you process.
Quickstart
One base_url. That is the whole migration.
Use the SDK you already have, or call the REST endpoint directly.
transcribe.py
from openai import OpenAIclient = OpenAI( base_url="https://api.inferentium.ai/v1", # the only change api_key="tapi_...",)tx = client.audio.transcriptions.create( model="inferentium-stt-1", file=open("call.mp3", "rb"),)print(tx.text)Comparison
Priced like a commodity.
The same drop-in interface, at a fraction of the cost.
| Inferentium | Whisper API | Typical managed ASR | |
|---|---|---|---|
| Price per hour of audio | ~$0.05 | $0.36 | $0.25 – $1.44 |
| Speed on long audio (realtime factor) | 230× | ~10–20× | 1–15× |
| OpenAI-compatible API | Yes | Native | Rarely |
| Billing | Prepaid, per second | Per minute, postpaid | Subscription / commit |
| Async jobs + webhooks | Included | No | Varies |
Based on public list prices as of 2026. Actual throughput varies with audio length and concurrency.
Accuracy
Measured in en-US, not promised.
Word error rate on FLEURS en-US (test), against the reference transcripts that corpus ships — same clips, same normaliser, every run.
2.1%
word error rate in en-US, on 5-10 s of audio
0.94 s
end to end for 5-10 s of audio, measured from another continent
0.5%
character error rate on the same audio
See the full spec sheet →Measured from São Paulo against us-central1, Opus 48 kbps.
Your audio is not training data
We use your audio only to produce your transcript. Nothing is retained to train models.
Prepaid, no lock-in
No subscription and no minimum commitment. Spend credits as you go and stop whenever you want.
Where your data is processed
Everything runs in Google Cloud region us-central1 (United States) today. Regional deployment is on the roadmap — talk to us if you have specific requirements.
Pricing
Commodity pricing. No lock-in.
Billed per second of audio via prepaid credits. Estimate: 1,000 hours of pre-recorded audio ≈ $50.
Pay as you go
$0.05 / hour of audio
- OpenAI-compatible API
- 25 languages, auto-detected
- Sync & async modes
- Prepaid credits, no minimums
Scale
Volume discounts
- Everything in Pay as you go
- Lower per-hour rate at volume
- Higher rate limits & priority
- Usage analytics & alerts
Enterprise
Custom
- Dedicated GPU fleet
- VPC / on-prem deployment
- Custom contract and billing
- SLA & support
Rates per product
One model per job, and every job priced on its own. Bulk throughput is cheap; an answer inside the same request, a live socket or a whole conversation takes a GPU slot the moment you ask for it.
| Product | Price | What it is |
|---|---|---|
| Pre-recorded | $0.05 per hour of audio | Upload a file, or submit a job and get a webhook. No duration limit. |
| Sync | $0.12 per hour of audio | One request, transcript in the response. Up to 2 minutes — dictation, commands, voice notes. |
| Realtime | $0.12 per hour of audio | A WebSocket: partial text while the speaker talks, a final for each utterance. |
| Voice agent | $1.50 per hour of call | A whole conversation on one socket — hearing, thinking and speaking. Flat per second of call, everything included. |
| Text to speech | $5.00 per million characters | A sentence in, audio out — mp3, Opus, WAV or raw PCM. Billed on the characters you send, which is the number you know before you call. |
| Speaker labelsAdd-on | +$0.02 per hour of audio | Who said what. Priced on top of a transcript, not instead of it. |
Billed per second of audio, or per character sent, from prepaid credits and rounded up per request. Failed requests are released and never charged.
FAQ
Questions, answered.
Ship audio intelligence today.
Get an API key in seconds, with 100 free credits. Change one line and you are live.
Get your API key →