Skip to content

Speech models for your vocabulary

Product names, drug doses and call signs are easy to mishear. We can adapt a model to your recordings and evaluate it inside your own network.

Which words does your model miss?

Choose an example to see how we’d score names and numbers. The error rates are illustrative. A 48-hour training sprint starts once we agree on recordings, evaluation and hardware.

Evaluation vocabulary

Medication names, doses and units

metoprolol succinate · 50 mg

Illustrative entity WER

Baseline16%
Sprint target8%

Scenario reduction of 8 percentage points. A held-out evaluation measures the actual change.

Estimate your GPU bill

Compare your API bill with the cost of running Oruk-1.7B yourself. Use your own GPU price and throughput, including idle time.

The starting values are examples. Storage, networking, operations, engineering and licensing cost extra.

Closed API / mo
$6,000
VPC compute / mo
$3,333
Compute savings
$2,667

Security and deployment terms

Audio retention

The public API discards raw audio after inference and keeps usage metadata. Optional diarization has separate retention terms. We can review zero-retention requirements for a private deployment.

Private networks

Our engineers review your hardware, container, network access and update process before agreeing to an air-gapped deployment.

HIPAA and SOC 2

Oruk currently has no HIPAA BAA or SOC 2 report. Keep protected health information out of the public API.

Licensing

The public ASR v2 weights use Apache-2.0. Before training, we review source licenses and your rights to the recordings.

Tell us about your audio

What do you record, how much do you process, and where does the model need to run?

Describe the vocabulary and recording conditions. Keep patient, customer and other sensitive details out of this form.