Skip to content
DUETA
Sign upConsoleLogin
Changelog

What's new

Notable, user-visible changes to the API and console. Breaking changes are called out explicitly; if one affects you we'll also email account holders.

27 August 2026

Credit is dollars now, and one rate covers every job

  • A balance is money. It reads $12.50 rather than a count of credits, in the console, on every invoice and in every API response.
  • One rate for everything: $0.50 per minute of audio, whichever engine ran the job. It is prorated by the second with no minimum charge, and it is charged on the length of the recording you upload, which the job already measures. A job with several microphones is billed on its longest file, not on their sum.
  • The plans are gone. There is nothing to subscribe to and nothing to pick: you top up and you spend it. Credit never expires, and a job that fails or is canceled is refunded.
  • Every new account is granted $20 of credit on signup, which is 40 minutes of audio.
  • Top-ups take any whole dollar amount from $5 to $500, with one-click amounts at 10, 20, 50 and 100. Purchasing is still not open while the payment provider integration is finished, so the endpoints answer the shape below and accounts run on their signup credit in the meantime.
API
GET /v1/credits returns balance_usd, and each ledger row carries delta_usd
GET /v1/usage returns billed_usd per day alongside billed_seconds
jobs carry estimated_cost_usd and billed_usd
GET /v1/billing/topup replaces GET /v1/billing/packs
POST /v1/billing/checkout takes amount_usd; orders carry topup_usd and credit_usd
22 August 2026

Enhancement defaults off when vocal separation runs

  • Denoising is no longer part of a default run when vocal separation is on. On recordings the separator already handles well it made no audible difference, and dropping it takes a stage out of every job.
  • Asking for denoising explicitly still runs it, and it is still on by default when vocal separation is off. Nothing you ask for changes meaning.
  • On difficult material — heavy crosstalk, or two people talking across each other for much of the take — denoising is still worth up to about 2 dB. It softens a poor result rather than rescuing one, so ask for it when your recording looks like that.
API
enhancement now defaults to false when vocal_separation is true
21 August 2026

Enhancement now runs after the speaker split

  • Denoising now runs on each separated speaker track rather than on the still-mixed recording. The denoiser only understands one voice at a time, and on a mixture it could mistake the overlapped second speaker for noise and suppress it. Separation quality on overlapped speech improves markedly.
  • A job reports its phases in the new order, so follow the phase by name rather than by its position in the run.
  • Track names record the new order too. Read the names a finished job gives you instead of composing them yourself.
  • Nothing about how you submit a job changed: same request, same options, same defaults, same billing.
API
order: vocal_separation → separation → enhancement → super_resolution → si_sdr
conversation-00_vocals_spk1_enh_48k (was …_vocals_enh_spk1_48k)
21 August 2026

Super-resolution now delivers 48 kHz

  • Upsampled tracks now come back at 48 kHz instead of 24 kHz. The model had been reconstructing the full band all along and the result was being downsampled on the way out; it no longer is.
  • Files are correspondingly about twice the size. Billing is unchanged — one charge for the length of the recording you upload.
  • The recording a job was made from can now be downloaded back, and the console shows it above the result tracks. It is kept for the same window as the separated tracks.
API
result.sample_rate reads 48000; the stem suffix is _48k
GET /v1/jobs/{id}/inputs/{name} returns the original recording
21 August 2026

Choose which phases run

  • You can now switch off vocal separation, denoising, upsampling, or the quality score. Everything is on unless you say otherwise, so an existing integration is unchanged.
  • Every job answers with the configuration the server actually resolved, so you can check what will run rather than assume it. A value we cannot read is rejected by name instead of quietly falling back to a default.
  • Speaker separation is not on the list: it is the job itself, not an option.
  • A job reports progress across only the phases it will actually walk through. With upsampling off the output stays at its native rate, and with scoring off no quality figure is returned.
  • In the console, the separation workbench carries the same four switches on its upload card.
  • Turning phases off shortens the run; it does not reduce the charge. A job is still billed for the length of the recording you upload.
API
vocal_separation · enhancement · super_resolution · scoring (all default true)
19 August 2026

One request for the whole pipeline

  • Breaking: there is now one way to create a job. Send one recording and the whole pipeline runs server-side inside it.
  • Running a single stage on its own, and chaining one job onto another, are both gone. Anything still calling for them will be turned away.
  • Billing is once per job, for the length of the recording you upload, instead of once per stage.
  • A running job names the phase it is on and reports progress across the whole pipeline, and a finished one carries the two speaker tracks, a quality score on each, and the rate they were written at.
API
POST /v1/jobs/separation — the only job-creation route
the per-stage routes and source_job_id now answer 404
19 August 2026

Key rotation, and stricter upload checks

  • An API key can be regenerated in place, from the console or over the API. The new key is shown once and the old one stops working immediately.
  • Upsampling accepts two tracks in one job, so a separated pair can be raised together. Both must be the same length, and a mismatch is refused at submission instead of failing inside the model.
  • Uploads that are too short, silent, or malformed are refused immediately with a specific message, rather than failing minutes later as a dead job.
  • Every job quotes its price before the work starts, so you can show a cost up front.
API
POST /v1/keys/{id}/rotate
estimated_cost_seconds on every job
18 August 2026

Five stage APIs, and chaining

  • Initial release. Vocal separation, denoising, speaker separation, upsampling, and quality scoring are each callable on their own.
  • Speaker separation takes one recording of two overlapping speakers and returns a track per speaker.
  • Any stage can be chained onto another, and intermediate audio never leaves the server. Each stage is billed separately.
  • A running job can be cancelled. Queued jobs stop at once, running jobs stop at their next checkpoint, and neither is billed.
  • Console: credit ledger, per-day usage, job history, and API key management. The separation workbench runs a job end to end from the browser.

The documentation always describes current behaviour; this page is only the history of how it got there.