Why did we open-source our inference engine? Read the post

Find the part number someone said in a noisy meeting

Superlinked gives your agent one API to turn call, meeting and hearing recordings into transcripts it can search and quote.

Get started
Meeting room, table microphone AMI meeting ES2004c
Human transcript

Um, findings, I've got my nice little picture there. This is uh the chip called the TA double one eight three five which is what's used in pretty much every remote control 'cause it sends out standard signals, based on your input and um it's pretty much used for all all TV 'Kay. remotes at the moment. Um and then we we're loo looking into battery options.

Whisper transcript

I've got my nice little picture there. This is the chip called the TA11835, which is what's used in pretty much every remote control because it sends out standard signals based on your input. And it's pretty much used for all TV remotes at the moment. And then we're looking into battery options.

A search finds 56 of the 61 spoken names, prices and part numbers

View evidence
Recorded telehealth consultation Urinary infection consultation

Clinician

Whisper transcript

Something called either nitrofurantoin or trimethoprine. And it's normally a three-day course. So one tablet twice a day for three days.

4.3% word error rate against 23 human-transcribed words

Human transcript Whisper
Nitrofurantoin nitrofurantoin
Trimethoprim trimethoprine Missed
three day course three-day course
one tablet, twice a day one tablet twice a day
three days three days
PriMock57 day 1, consultation 09, CC BY 4.0
Supreme Court argument Commissioner of Internal Revenue v. Zuch

Shay Dvoretzky, counsel for the taxpayer

Whisper transcript

There is a recent report from the IRS Taxpayer Advocate showing that between 2004 and 2018, only a little over 1% of levy notices actually resulted in a CDP hearing being sought.

20.0% word error rate against 40 human-transcribed words

Human transcript Whisper
IRS Taxpayer Advocate IRS Taxpayer Advocate
2004 2004
2018 2018
1 percent 1%
levy notices levy notices
CDP hearing CDP hearing
Commissioner of Internal Revenue v. Zuch, No. 24-416, transcript page 47, line 10
Meeting room, table microphone Remote control project kickoff

Project manager (speaker B), with room backchannel

Whisper transcript

So according to the group, we're going to be selling this remote control for €25 and we're aiming to make €50 million. So we're going to be selling this on an international scale and we don't want it to cost any more than €1250, so 50% of the selling price.

3.7% word error rate against 54 human-transcribed words

Human transcript Whisper
twenty five Euro €25
fifty million Euro €50 million
international scale international scale
twelve fifty Euros €1250 Missed
fifty percent 50%

“€1250” is a hundred times the spoken €12.50, which the speaker calls fifty percent of the €25 price.

AMI meeting ES2002a, CC BY 4.0

3 of the 12 recorded clips play here, with two more in the hero and the playground. Across all 12, Whisper's pooled word error rate is 8.1% over 594 human-transcribed words, and 5 key terms came back wrong, 2 of them shown above.

One call turns a recording into text your agent can search

View on GitHub
from pathlib import Path
from sie_sdk import SIEClient
client = SIEClient(
api_key="API keysk-sie-…",
base_url="https://api.superlinked.com",
)
audio = {"data": Path(fileMP3scotus-nrc-atomic-energy-act.mp3browse).read_bytes(), "format": "mp3"}
result = client.extract(
"modelopenai/whisper-large-v3-turbo",
{"audio": audio},
instruction="Transcribe the audio verbatim.",
)
print(result["data"])
audio_bytes=$(base64 < 'scotus-nrc-atomic-energy-act.mp3' | tr -d '\n')
curl https://api.superlinked.com/v1/extract/openai%2Fwhisper-large-v3-turbo \
  -H "Authorization: Bearer sk-sie-…" \
  -H "Content-Type: application/json" \
  -d "{\"items\":[{\"audio\":{\"data\":\"$audio_bytes\",\"format\":\"mp3\"}}],\"params\":{\"instruction\":\"Transcribe the audio verbatim.\"}}"
Build the "Speech to text" capability into my app using the Superlinked Inference Engine (SIE).

Context
- SIE is an OpenAI-style inference API. Python SDK: `from sie_sdk import SIEClient`; TypeScript: `@superlinked/sie-sdk`.
- Base URL: https://api.superlinked.com (or my regional endpoint). Auth: Bearer key from env `SIE_API_KEY` (never hard-code it).
- Model: openai/whisper-large-v3-turbo (SIE primitive: /extract). Keep the model id configurable.

Task
- Input: an uploaded audio recording.
- Behaviour: return a verbatim transcript of the uploaded audio
- Call the selected SIE primitive once per request and map the response into your domain type.

Deliverables
- A typed client wrapper, an application-level function for this task, error handling for timeouts/empty input, and unit tests with a stubbed client.
- Wire it into my existing stack (ask me which framework if unclear) and add a short usage example.
Output
source · scotus-nrc-atomic-energy-act.mp3
text
First, I would point the Court to the licensing provisions in the Atomic Energy Act, which are 42 U.S.C., 2073A, which deals with special nuclear material, 2093, which deals with source material, and 2111, which deals with byproduct material.
duration_ms
18111
language
null
whisper-large-v3-turbo

Speech to text quality and latency

PRICE
$ / audio minute
$0.0075 AssemblyAI Universal-3 Pro streaming
$0.0077 Deepgram Nova-3 streaming
$0.0139 SIE Whisper Large v3 Turbo
Google Speech-to-Text v2 $0.016
AWS Transcribe $0.024
QUALITY
1 − WER
SIE Whisper Large v3 Turbo 0.89
0.87
0.86
0.85
0.84
LATENCY
p50 ms
SIE Whisper Large v3 Turbo 350ms
380ms
390ms
410ms
430ms

Deploy your way

Self-host with K8s

Easy & scalable deployment in your own cloud.

  • Terraform to your cloud in minutes
  • Apache-2.0, same engine as Cloud
  • Scales to zero, no bill between jobs
  • Per-tenant pools, no noisy neighbors
Agent prompt
Deploy SIE to our AWS account with the superlinked/sie/aws Terraform module. Docs: superlinked.com/docs/deploymentDeploy SIE to our GCP project with the superlinked/sie/google Terraform module. Docs: superlinked.com/docs/deploymentDeploy SIE to our Azure AKS cluster via helm install. Requirements: superlinked.com/docs/deployment
Deploy guide

Run locally

Run the same models on your own machine.

  • Runs on NVIDIA GPU or Apple Silicon
  • One command, no Docker or cluster
  • All 100+ Cloud models, fully offline
  • Same SDK and IDs, no code changes
pip install "sie-server[local]" && sie-server servepip install "sie-server[local]" && sie-server serve --device cuda
Quickstart

Contact us

Tell us about your use case and we'll get back to you shortly.

Apply for an inference grant

Free capacity on our hosted cluster for selected projects. Tell us what you run and we reply by email.