New features, improvements and fixes to the API, the engine and the platform, as they ship.
With who said what turned on, each sentence now stays with the person who said it. The last word of a sentence no longer jumps to the next speaker, and short replies like "Sure." or "Of course." stay on the right side of the conversation.
If your client disconnects or times out before a POST /v1/audio/transcriptions request starts processing, Orchard cancels it: your concurrency slot frees up and you're not billed for it.
Drop a call, a meeting or a voice note at orchardrun.com/studio and watch the transcript appear as it's processed. Turn on who said what, rename the speakers, search the text, click any line to jump to it in the audio, and download it as TXT, SRT or VTT. Long files are never cut off: leave the tab open or come back later with the same link.
Canceled your plan and want a different one? You can pick it right away, no need to wait for the current period to end. Your remaining minutes carry over, and the new plan starts its own monthly cycle the day you pay.
POST /v1/audio/transcriptions now takes the fast lane for short clips: a 12-second clip comes back in about 2.5 seconds of processing. We also documented prompt, which lets you pass vocabulary and context (product names, medical terms) to improve accuracy on your domain.
Upgrading from the billing portal now updates your account immediately, and the extra minutes of the bigger plan are credited right away instead of at the next renewal.
Transcript text is no longer sent to any third-party AI service. Every step, from transcription to speaker labels, runs on infrastructure Orchard operates, by default. We also stopped logging full request URLs, so tokens in your webhook URLs never end up in logs.
We moved production to a new platform built for reliability: automatic database backups every 15 minutes, services that restart themselves on failure, and deploys that take minutes and can be rolled back with one command.
The full Orchard stack, transcription and diarization, now runs inside your own servers. Audio never leaves your network, nothing calls home, and the license works offline. Built for banks, healthcare and government. Talk to us.
Short audio skips the queue entirely and is transcribed inside your request. For dictation, voice notes and chunked recordings, waiting time drops to almost nothing.
Transcribe from Claude, Cursor or any MCP-compatible assistant: ask for a transcript in plain language and your assistant does the rest. Four tools: transcribe_file, transcribe_url, get_transcription and get_usage. Install with npx orchard-mcp.
If you don't send language, Orchard detects it instead of assuming English.
We ran seven speech-to-text APIs on the same 13 real Latin American calls and published everything: the audio, the code and the results. Orchard came fourth, at a fraction of the price of the three ahead. See the benchmark.
Zorzal-1 got more robust on long recordings with background noise and silence: fewer repeated or dropped passages.
If a diarization backend stops responding, Orchard switches to the next one automatically. You keep getting speaker labels, with no errors on your side.
diarize=true now works on POST /v1/audio/transcriptions, the same endpoint you already use with the OpenAI SDK. In the playground you can also tell Orchard how many speakers to expect.
Every page on orchardrun.com has a Markdown twin, plus llms.txt and llms-full.txt. Point your assistant at the docs and it gets clean text instead of HTML.
When an upload looks degraded, the response tells you, so you know why a transcript came out worse before you blame the model.
A new pricing page shows the real cost per hour of audio next to the main alternatives, with a calculator for your monthly volume. See pricing.
Words are now assigned to speakers one by one instead of by segment, so fast back-and-forth exchanges keep the right speaker on each word.
Need more than Pro? Choose exactly how many hours you need each month and pay the same flat per-minute rate, straight from the billing page.
Know who said what on every plan, with a monthly diarization quota per tier. The playground now accepts file uploads, so you can try it on your own recordings.
Share your referral link and earn bonus minutes when the people you invite subscribe.
The documentation no longer requires an account. New accounts are confirmed with an email code.
An official TypeScript SDK, and a provider for the Vercel AI SDK so Orchard plugs into apps you already build with it.
Orchard has been shipping since well before this page existed. Earlier releases, from the first version of the engine onward, live in our archive.
