選擇符合你技術棧的 SDK
提交任務、輪詢一次、拿到 6 條 stem
12 行程式碼就能證明介面範圍。提交一次分離，取得 job id，輪詢一次，收到人聲、鼓、貝斯、吉他、鋼琴和 other——每個 SDK 都跑同一套 htdemucs_ft 分離流程。依輸入音訊分鐘數從永不到期的分鐘包扣費（依購買量不同，每分鐘 $0.08–$0.14）。
# 1) Submit a split job
curl -X POST https://api.aistemsplitter.org/v1/audio/splits \
-H "Authorization: Bearer $AISTEMSPLITTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"input": { "type": "direct_url", "url": "https://example.com/song.mp3" },
"stemModel": "htdemucs_ft"
}'
# → { "id": "split_01J…", "status": "queued" }
# 2) Poll for completion
curl https://api.aistemsplitter.org/v1/audio/splits/split_01J… \
-H "Authorization: Bearer $AISTEMSPLITTER_API_KEY"
# → { "status": "succeeded", "stems": { "vocals": "...", ...x6 } }同一個 API，也給團隊其他人一條無程式碼路徑
工程師很少是唯一使用者。把同一個 AI Stem Splitter API 放進 n8n、GitHub Actions、container 或 Pipedream workflow，讓營運、內容和 CI 都跑在同一個 vendor 上。
開發者安裝前常問的問題
Which SDK should I install — Node, Python, or curl?
Pick the language your service already runs on. Node and Python ship typed clients with one-call helpers (waitForSplit, verify_signature) and full method autocomplete; curl works on the same REST endpoints if you're prototyping or scripting from a shell. The hosted API is identical across all 13 distribution channels — same job ids, same webhooks, same per-minute billing.
Do credits expire if I don't use them?
No. Credit packs never expire — purchase $6.99 / 50 minutes, $15 / 150 minutes, or $39.99 / 500 minutes and the balance stays on your account until you spend it. There is no monthly subscription, no rollover cap, and no auto-renewal. The first 3 minutes are free with no card required, charged at $0.08–$0.14 per minute depending on pack size.
Which model do you use, and can I name it in the request?
The default is htdemucs_ft (Demucs fine-tuned, 4-stem: vocals, drums, bass, other) — the same model variant the open-source community benchmarks at ~8.7 dB SDR. You can switch to htdemucs_6s for guitar and piano stems by passing model='htdemucs_6s' on submit. Models are named explicitly in every SDK so output quality is comparable to local Demucs runs.
Can I self-host the SDK or the inference?
The SDK clients are MIT-licensed and ship to npm, JSR, PyPI, Maven Central, Packagist, LuaRocks, SPM, and CocoaPods — fork or vendor them as you like. Inference itself runs on our managed GPUs against the hosted API at api.aistemsplitter.com — there is no on-prem inference image in v1, but the GHCR Docker image wraps the SDK for self-hosted orchestration.
What's the rate limit, and how do I handle large files?
Concurrent jobs and per-minute request caps live on /developers/api with the current numbers (they age, so they aren't pinned here). For large uploads, call presignUpload (Node) / presign_upload (Python) to get a direct-to-storage URL — the SDK handles retries with exponential backoff out of the box. Files up to 50 MB upload directly; larger files use multipart presigned URLs.
How do I receive the result without polling forever?
Pass webhook_url on createSplit. When the job finishes, the API POSTs an HMAC-SHA256-signed event to your URL with the stem download links inline. Verify the signature with one SDK call (verifyWebhook in Node, verify_signature in Python). The Webhooks section in each per-language quickstart shows runnable Express, Hono, FastAPI, and Flask handlers.
我該安裝哪個 SDK——Node、Python 還是 curl？
選擇你的服務已經在用的語言。Node 和 Python 提供型別化 client，帶有一次呼叫的 helper（waitForSplit、verify_signature）和完整 method autocomplete；如果你只是從 shell 做原型或寫腳本，curl 也使用同一套 REST endpoints。託管 API 在全部 13 個分發渠道中都一致——同樣的 job ids、同樣的 webhooks、同樣的按分鐘計費。
分鐘額度不用會過期嗎？
不會。分鐘包永不到期——購買 $6.99 / 50 分鐘、$15 / 150 分鐘或 $39.99 / 500 分鐘後，餘額會留在你的帳戶直到用完。沒有月訂閱、沒有 rollover 上限，也不會自動續費。前 3 分鐘免費且不需信用卡，依方案大小不同，每分鐘 $0.08–$0.14。
你們使用哪個模型？我可以在 request 裡指定嗎？
預設是 htdemucs_ft（Demucs fine-tuned，4-stem：vocals、drums、bass、other）——這是開源社群 benchmark 約 8.7 dB SDR 的同一個模型變體。提交時傳入 model='htdemucs_6s'，即可切換到支援 guitar 和 piano stems 的模型。每個 SDK 都明確命名模型，所以輸出品質可和本機 Demucs 執行結果比較。
我可以 self-host SDK 或推論服務嗎？
SDK clients 採 MIT 授權，並發布到 npm、JSR、PyPI、Maven Central、Packagist、LuaRocks、SPM 和 CocoaPods——你可以依需要 fork 或 vendor。推論本身跑在我們管理的 GPU 上，透過 api.aistemsplitter.com 的 hosted API 呼叫——v1 沒有 on-prem 推論 image，但 GHCR Docker image 會包住 SDK，供 self-hosted orchestration 使用。
速率限制是多少？大檔案要怎麼處理？
並行任務數和每分鐘 request 上限以 /developers/api 的目前數值為準（這些數字會變，所以這裡不固定寫死）。上傳大檔案時，呼叫 presignUpload（Node）/ presign_upload（Python）取得 direct-to-storage URL——SDK 內建 exponential backoff retry。50 MB 以內檔案可直接上傳；更大的檔案使用 multipart presigned URLs。
要怎麼不用一直輪詢也能收到結果？
在 createSplit 傳入 webhook_url。任務完成後，API 會向你的 URL POST 一個 HMAC-SHA256 簽名事件，並在 payload 內直接帶上 stem 下載連結。用一次 SDK 呼叫驗證簽名（Node 裡 verifyWebhook，Python 裡 verify_signature）。每個語言 quickstart 的 Webhooks 區塊都提供可執行的 Express、Hono、FastAPI 和 Flask handlers。