AudioShake

Python client and command-line tool for the AudioShake API: stem separation, dialogue isolation, speech clean-up, multi-speaker separation, music detection and removal, lyric transcription and alignment.

pip install audioshake          # library + CLI
uvx audioshake run song.wav -m vocals,drums   # or run the CLI with no install

• Library: AudioShake and AsyncAudioShake clients, typed results, precise exceptions, automatic retries, one-call run() or the individual API steps.
• CLI: audioshake run for files, folders and URLs with a live progress tree, --json for scripts, audioshake models and audioshake tasks to explore your account.
• Python 3.10+, typed (py.typed), tested on 3.10 through 3.15.
• Version 0.x: the surface is settling; breaking changes are called out in the changelog and bump the minor version.

Authenticate either with an API key from the AudioShake dashboard, exported as AUDIOSHAKE_API_KEY, or by signing in from the terminal with audioshake login.

Library

from audioshake import AudioShake

with AudioShake() as client:  # reads AUDIOSHAKE_API_KEY
    result = client.run("song.wav", ["vocals", "drums"], output_dir="stems/song")

print(result.task.cost)  # credits charged
print([str(f) for f in result.files])  # stems/song/vocals.wav, stems/song/drums.wav

run() uploads the file (or passes a URL through), creates the task, waits for it and downloads the outputs into output_dir as <output>.<format>, for example vocals.wav. Give each source its own directory; the CLI does that for you with one sub-directory per source. Every step is also available on its own:

from audioshake import AudioShake, Target

with AudioShake() as client:
    for model in client.models.list():
        print(model.id, model.credits_per_minute, model.output_formats)

    asset = client.assets.upload("interview.mp3")
    task = client.tasks.create(
        asset,
        [
            Target("multi_voice", num_speakers=3),
            Target("transcription", formats=("json",), language="en"),
        ],
    )
    task = client.tasks.wait(task.id, on_poll=lambda t: print(t.done))
    client.tasks.download(task, "out/")

The async client has the same surface:

import asyncio
from audioshake import AsyncAudioShake


async def main():
    async with AsyncAudioShake() as client:
        results = await asyncio.gather(
            client.run("a.wav", ["vocals"], output_dir="out/a"),
            client.run("b.wav", ["vocals"], output_dir="out/b"),
        )


asyncio.run(main())

Errors are typed, so you can catch what you can handle:

from audioshake import AudioShake, AuthenticationError, TaskFailedError

try:
    with AudioShake() as client:
        result = client.run("song.wav", ["vocals", "drums"], output_dir="stems/song")
    for target in result.failed:  # some models failed; the rest were downloaded
        print(target.model, target.error)
except TaskFailedError as e:  # every model failed
    print(e.task.failed[0].error)  # the API's reason
except AuthenticationError:
    ...

Connection errors, timeouts, rate limits and 5xx responses are retried with backoff before an exception reaches you. Task creation is billed, so it is only retried when the API declined to act (408, 425, 429, 503). The library logs to the audioshake logger and never prints.

Full reference: docs/library.md.

CLI

audioshake run song.wav -m vocals,drums -o stems/
audioshake run ./album -m effects,dialogue,music_fx -f mp3 --residual
audioshake run https://example.com/talk.mp3 -m transcription --language en
audioshake run panel.wav -m multi_voice --num-speakers 3

audioshake models                 # what your account can run, with cost and formats
audioshake tasks list             # recent tasks
audioshake tasks get TASK_ID
audioshake tasks download TASK_ID -o out/

run takes any number of files, directories (-r to recurse) and public URLs, processes them concurrently (-c, default 10), and writes OUT/<source>/<output>.<format>, the layout Demucs and Spleeter use. In a terminal it paints a live tree with one row per source and per model; elsewhere, or with --plain, it logs lines instead.

Logs go to stderr; results go to stdout. --json makes the results machine-readable (run prints one JSON object per source):

audioshake --json run song.wav -m vocals | jq .files
audioshake --json models | jq '.[].id'

Exit codes: 0 success, 1 a source failed or partly completed, 2 usage error, 3 authentication, 4 network failure after retries, 5 unexpected error, 130 interrupted.

Full reference, every option and the output layout: docs/cli.md.

Install

Library	pip install audioshake or uv add audioshake
CLI, no install	uvx audioshake ...
CLI, on your PATH	uv tool install audioshake or pipx install audioshake
From source	git clone https://github.com/audioshake/audioshake-python && cd audioshake-python && uv sync

Authentication

Two credentials work everywhere; the first one found wins:

1. An API key: --api-key, or AUDIOSHAKE_API_KEY in the environment (what CI and scripts should use).
2. A browser sign-in session: audioshake login opens the AudioShake sign-in page, caches the session in ~/.audioshake/credentials.json (mode 600) and refreshes it silently until its refresh token expires.

audioshake login                # opens the browser
audioshake login --no-browser   # prints the URL (on stderr) instead
audioshake whoami               # which credential is in use, and who you are
audioshake logout

Under the hood login is the OAuth 2.0 authorization-code flow with PKCE: the CLI listens on 127.0.0.1 for the redirect, exchanges the code for tokens, and sends them to the API as Authorization: Bearer. From Python, use audioshake.auth.login() and pass the token to AudioShake(bearer_token=...).

AUDIOSHAKE_BASE_URL points the library at another API host.

Models

Model ids are passed through to the API, which is the source of truth for what exists; audioshake models lists the current catalogue with cost per minute, output formats, supported languages and input limits. Text models (transcription, alignment, music_detection, music_identification) produce JSON by default; audio models produce WAV unless you ask for another format. Several models run from one upload: -m effects,dialogue,music_fx gives the dialogue, music and effects stems, -m vocals_backing,vocals_lead the two vocal stems.

Claude Code and Codex plugin

The plugin/ folder is a plugin for Claude Code and Codex that teaches the agent to use the CLI and library. Install it from the AudioShake marketplace.

Inside Claude Code:

/plugin marketplace add audioshake/plugins
/plugin install audioshake-cli@audioshake

From your shell, for Codex:

codex plugin marketplace add audioshake/plugins
codex plugin add audioshake-cli@audioshake

Then ask for stems, an instrumental, isolated dialogue or lyrics from a file and the agent runs audioshake for you. The skill is kept in step with the CLI by tests/test_skill_docs.py.

Development

uv sync                          # environment with dev dependencies
uv run pytest                    # tests (HTTP is mocked; no key needed)
uv run ruff check . && uv run ruff format --check .
uv run mypy                      # strict on the library

CI runs the same on Python 3.10 to 3.15. Releases are published to PyPI from v* tags. See CONTRIBUTING.md before opening an issue or a pull request. Design notes are in docs/design.md.

Links

• API documentation
• Issues
• MIT License, © AudioShake, Inc.