Inferentium.AI
Get API key
Spec sheet

Our models, measured.

What each model does, what it costs you in latency, and how accurate it is — per language, from benchmarks we re-run instead of numbers we repeat.

inferentium-stt-1

Pre-recorded

Our own speech-to-text pipeline for files: one request in, punctuated and cased text out. Auto-detects the spoken language, batches on our GPUs, and bills per second of audio.

Languages
25
Audio in
wav · mp3 · m4a · flac · ogg/opus · webm
Sample rate
any (resampled to 16 kHz)
Max upload
500 MB
Returns
text · language · duration
Word timestamps
roadmap
Speaker labels
up to 4 speakers (add-on)
Audio retention
none
Endpoints
  • POST /v1/audio/transcriptions
  • POST /v2/transcribe (Sync, ≤ 2 min)
  • POST /api/v1/stt/transcribe/stream (SSE)
  • POST /api/v1/stt/jobs (async + webhook)
  • POST /v2/transcript (AssemblyAI-compatible)

inferentium-stt-1-realtime

Streaming

The same pipeline driven at live pace over a WebSocket: partial text while the speaker is still talking, a final for each utterance when they stop.

Languages
25
Audio in
PCM16 mono, 16 kHz frames
Partials
every 400 ms
Segmentation
VAD, per utterance
Returns
partial · final · end
Billed
per second of audio received
Audio retention
none
Endpoints
  • WSS /ws/realtime (short-lived upload token)
Accuracy

Word error rate, per language.

Accuracy is a property of a language, not of a logo. You are reading this page in en-US, so that is the table we put first.

en-US

Your languageFLEURS en-US (test) · 2026-09-11
Audio lengthWERCERp50p95Speed
Short clip (5-10 s)2.1%0.5%0.94 s1.28 s6×
Call segment (50-90 s)——1.92 s1.98 s37×
Meeting (9-11 min)——5.15 s5.44 s120×

pt-BR

FLEURS pt-BR (test) · 2026-09-10
Audio lengthWERCERp50p95Speed
Short clip (5-10 s)4.8%0.8%0.77 s0.8 s9×
Call segment (50-90 s)3.2%0.9%1.43 s1.66 s49×
Meeting (9-11 min)9.4%6.6%4.82 s5.46 s124×

A dash means we have not published that number yet. We are re-measuring English above 60 seconds after a change to how long audio is segmented; the latency and throughput in the same row are unaffected. We would rather show a dash than a number we do not trust.

Streaming

How live feels.

Measured over a real WebSocket session, audio pushed at live pace in 100 ms frames from São Paulo.

MetricMedianWhat it means
First word776 msFrom the start of speech to the first text on screen
Partial lag240 msHow far behind the speaker the running text stays
Final lag837 msFrom the end of an utterance to its final, corrected text
Final lag (p95)965 msWorst case in 19 of 20 utterances
WER4.9%On the same audio as the pre-recorded table

FLEURS pt-BR (test), measured 2026-09-10. Billed seconds are the audio you send plus a 1.5 s tail.

Method

How every number here was produced.

  • Public benchmark audio (FLEURS) with the reference transcripts it ships, concatenated into three duration classes so the same text is scored at 5 seconds and at 10 minutes.
  • Measured from São Paulo against our fleet in us-central1, so the latency includes the internet — about 130 ms of round trip that a customer in the US does not pay.
  • Uploads are Opus at 48 kbps in audio mode, what a good client sends. Lower bitrates and voip mode cost 2-3 points of WER on long files.
  • Error rates are the mean over every run and latency the median, on text normalised the same way for every provider we compare against.
  • We publish what we measured, including where it is worse than we would like. Numbers move when the pipeline moves.
Start free →Read the docsAsk about your audio