You do not need a key to screen a recording. The service answers a request that carries no credential at all, which is what someone using Swan through a platform or an assistant will do. Screening without a key is capped at 20 requests per hour, and a key raises that to 600, because screening is the route that runs the model.
A key is optional, and it is issued to this browser as soon as you ask for one. It gives your own client its own hourly allowance instead of sharing whatever else arrives from your address, and it is what lets you look up an earlier result by its id. There is no account to create, no email to confirm, and nothing to pay. Only the hash of a key is stored, so the key itself is shown once and cannot be looked up afterwards.
Two keys are issued per address per hour and six per day, with a daily ceiling for the service as a whole. When a limit is reached the response says how long to wait. It does not fail silently, and it does not hand out a key that will not work.
Screening is available with or without a key. A key gets its own hourly allowance, separate from anything else that shares your address, and is what lets you look up an earlier analysis by its id. There is no account to manage, no plan to select, and no other data the service will return.
Send a recording and receive three results: whether the voice is likely human or likely synthetic, the quality of the evidence behind that finding, and one recommended action, which is proceed, verify or hold.
At least ten seconds of speech for a finding, up to 120 seconds per request, and 25 MB. A shorter recording returns an inconclusive result rather than an error, and a longer one is refused with the reason stated.
Wav, mp3, m4a, ogg, flac and webm. Telephone band and compressed audio are accepted, and reduce what the service is willing to report rather than producing a finding.
Not speaker identification, not a calibrated probability, and not evidence on its own. A result describes a recording. It never claims that the person speaking has been identified.
Audio is decoded in memory and never stored. A metadata record is kept for 30 days so a result can be explained later. Details are in the Swan privacy policy.
Send the key in an Authorization header as a bearer token. The recording goes in the file field, and the decision comes back in the same response.
Documentation: api.humsana.com/docs. OpenAPI document: api.humsana.com/openapi.json.
The service states what a recording supports, and says so when it supports neither finding.