在 Python 裡做 stem 分離。

Default to sync (StemSplitter) — that's what the 12-line hello world uses, and it's what fits naturally inside a Django view or a Celery task. Switch to AsyncStemSplitter when you're already on FastAPI, Starlette, or asyncio in a notebook: the method names match (await client.submit, await client.wait, await client.download), so swapping costs one import + one await per call.

Yes — both are first-class. The Webhooks section above ships runnable handlers for each: FastAPI uses async def webhook(request: Request) with verify_signature; Flask uses a sync @app.post route. Both verify HMAC-SHA256 with one import (from aistemsplitter import verify_signature) and unpack presigned stem URLs from the JSON body without extra parsing.

Standard tooling: `python -m venv .venv && source .venv/bin/activate && pip install aistemsplitter` for venv; `poetry add aistemsplitter` for Poetry; `uv add aistemsplitter` for uv. The package is pure-Python wheels (no native compilation), so installation completes in seconds on every platform — no pyenv shim, no FFmpeg pre-step, no GPU driver.

Yes. We expose the same model file the open-source community trusts — htdemucs_ft, plus htdemucs and htdemucs_6s — so output is bit-comparable to a local Demucs run with the same model + same input, modulo our managed inference settings (batch size, precision). Pass model='htdemucs_6s' on submit when you need guitar and piano stems.

Drop straight to /developers/api. The SDK is a thin typed wrapper over the same REST endpoints the curl quickstart calls, so the auth header (Authorization: Bearer ast_live_…), the webhook signature scheme (HMAC-SHA256 in aistemsplitter-signature), and the response shapes are identical. You can mix-and-match — call submit via the SDK, fetch status via raw httpx for a custom retry policy.

Two strategies. (1) Use webhooks — set webhook_url on submit and skip the polling loop entirely; the API posts results when the job finishes regardless of duration. (2) For files larger than 50 MB, call client.presign_upload() to get a direct-to-storage URL, upload via httpx multipart, then submit with the returned audio_url instead of streaming through our gateway.

預設使用 sync（StemSplitter）——12 行 hello world 用的就是它，也自然適合 Django view 或 Celery task。當你已經在 FastAPI、Starlette 或 notebook 裡的 asyncio 環境中時，再切換到 AsyncStemSplitter：method names 保持一致（await client.submit、await client.wait、await client.download），所以替換成本只有一個 import 加每次呼叫一個 await。

可以——兩者都是 first-class。上方 Webhooks 區塊為每個框架提供可執行 handler：FastAPI 使用 async def webhook(request: Request) 搭配 verify_signature；Flask 使用 sync @app.post route。兩者都透過一個 import（from aistemsplitter import verify_signature）驗證 HMAC-SHA256，並從 JSON body 解出 presigned stem URLs，不需額外解析。

使用標準工具即可：venv 用 `python -m venv .venv && source .venv/bin/activate && pip install aistemsplitter`；Poetry 用 `poetry add aistemsplitter`；uv 用 `uv add aistemsplitter`。這個 package 是 pure-Python wheels（不需 native compilation），所以每個平台上都能幾秒完成安裝——不需要 pyenv shim、不需要 FFmpeg 預步驟，也不需要 GPU driver。

會。我們暴露的是開源社群信任的同一個模型檔——htdemucs_ft，以及 htdemucs 和 htdemucs_6s——所以在相同模型和相同輸入下，輸出可以與本機 Demucs 執行結果做 bit-comparable 對比，差異只來自我們的 managed inference settings（batch size、precision）。當你需要 guitar 和 piano stems 時，在 submit 傳 model='htdemucs_6s'。

直接切到 /developers/api。SDK 只是同一套 REST endpoints 上的 thin typed wrapper，也就是 curl quickstart 呼叫的同一套介面，所以 auth header（Authorization: Bearer ast_live_…）、webhook signature scheme（aistemsplitter-signature 中的 HMAC-SHA256）和 response shapes 都完全一致。你可以混用——用 SDK submit，用 raw httpx fetch status 來實作自訂 retry policy。

兩種策略。1）使用 webhooks——submit 時設定 webhook_url，完全跳過輪詢迴圈；不論時長多久，API 都會在 job 完成後 POST 結果。2）對於大於 50 MB 的檔案，呼叫 client.presign_upload() 取得 direct-to-storage URL，透過 httpx multipart 上傳，然後用回傳的 audio_url submit，而不是把檔案串流經過我們的 gateway。