mixing

Tools for video and audio editing in Python — slicing, fades, audio replace/mix, Ken Burns, thumbnails, subtitles, speech-to-text + filler removal, TTS dubbing, and chapter detection.

pip install 'mixing[media]'   # audio + video editing (moviepy, opencv, scipy, pydub)
pip install mixing            # base: only the pure modules (srt, chapters, errors, util, egress, transcript, TTS)

The base install is light on purpose: it carries no media backend (and so no bundled ffmpeg binary). mixing.audio and mixing.video need the media extra; importing them without it raises an ImportError that says so.

---

🤖 AI-first: start with the skills & agents

mixing is built to be driven by AI coding agents. The fastest, most reliable way to use it is through the bundled skills and agents in .claude/ — task-oriented playbooks that encode the right functions, the calling conventions, and the gotchas, so an agent gets it right the first time.

Skills (.claude/skills/) — how to use mixing for a task:

Skill	Use it for
mixing	Start here — router, architecture, the output protocol, extras
mixing-audio	slice, fade, crop, concat, overlay, normalize, align, segment audio
mixing-video	slice, crop, loop, speed, replace/normalize audio, Ken Burns, thumbnails, subtitles
mixing-transcript	transcribe, remove fillers, SRT/prose, chapters
mixing-dubbing	TTS re-voicing & SRT translation

Agents (.claude/agents/):

• mixing-editor — performs an editing task end-to-end ("edit this media …").
• mixing-dev — works on the mixing codebase.

Using Claude Code? These are discovered automatically from .claude/. The architectural contract they share lives in .claude/CLAUDE.md.

---

Quick start

import mixing  # cheap: no moviepy/opencv loaded yet

assert mixing.has_ffmpeg()  # most ops need ffmpeg on PATH

# --- audio ---
from mixing.audio import Audio

Audio("song.mp3")[10:30].fade_in(2).fade_out(3).save(output="clip.mp3")

# --- video ---
from mixing.video import replace_audio, ken_burns_video

replace_audio("clip.mp4", "music.mp3", mix_ratio=0.7, output="mixed.mp4")
ken_burns_video("cover.jpg", duration=8, output="cover.mp4")

# --- transcript: remove "uh"/"um" (ElevenLabs Scribe + ffmpeg) ---
from mixing.transcript import remove_fillers

result = remove_fillers("talk.mov", "out/")  # uses $ELEVENLABS_API_KEY
print(result.cleaned_media)

Core ideas (learn once)

The output protocol. Every result-producing function takes one output argument whose role is constant and type is open:

crop_video("in.mp4", 5, 15, output="out.mp4")  # file → writes it, returns Path
crop_video("in.mp4", 5, 15, output="clips/")  # dir  → auto-named file inside
crop_video("in.mp4", 5, 15)  # None → saves beside the input
fade_in("in.mp3", output=lambda audio: audio)  # callable → receives the result

Object producers (audio editing) return the in-memory object when output=None, so you can chain and save when ready. Functions that emit several artifacts qualify the destinations (remove_fillers(media, output_dir=…, output_media=…)).

Lazy & tiered. import mixing (and import mixing.chapters / mixing.srt) pull no heavy backends; moviepy/opencv/pydub load only when you touch a name that needs them. Import the facade (mixing.Video) or the subpackage (from mixing.video import Video).

Sliceable media. Audio and Video are lazy, sliceable, context-managed views:

from mixing.video import Video

with Video("movie.mp4") as v:
    v[10:30].save(output="cut.mp4")  # seconds; v[100] returns a single frame

Logging, not printing. mixing never writes progress messages to stdout — every module logs through logging.getLogger(__name__), and the package root carries a NullHandler, so mixing is silent by default. This keeps library calls usable from a stdout-parsing consumer. To see mixing's own progress messages (e.g. while working interactively, or when an operation like mixing.video.genai.generate_video fails and you want the full diagnostic), opt in with:

import logging

logging.basicConfig(level=logging.INFO)  # or level=logging.ERROR for errors only

Requirements

• ffmpeg on PATH (most operations). macOS: brew install ffmpeg; Debian/Ubuntu: sudo apt install ffmpeg; Windows: download from ffmpeg.org and add bin to PATH.
• ElevenLabs API key for mixing.transcript / mixing.dubbing (ELEVENLABS_API_KEY, or pass api_key=). Responses are cached on disk, so re-runs are free and offline.
• Google Cloud auth for AI video generation (mixing.video.genai, Vertex AI Veo).

Optional extras

pip install mixing[media]      # moviepy / opencv / scipy / pydub: mixing.audio + mixing.video
pip install mixing[audio]      # ffmpeg-python / soundfile (on top of media)
pip install mixing[widget]     # interactive Jupyter audio widget
pip install mixing[gen]        # Google Vertex AI Veo generation
pip install mixing[llm]        # aix — chapter titling + SRT translation
pip install mixing[clipboard]  # get file paths from the clipboard
pip install mixing[beats]      # librosa — beat/downbeat/onset analysis
pip install mixing[ml]         # torch + transformers — find_segments(strategy="audioset"): songs vs jingles vs talk

What's inside

Subpackage	Highlights
mixing.audio	Audio, fade_in/out, crop_audio, concatenate_audio, overlay_audio, loop_audio, duck_audio, find_audio_offset, find_segments/extract_segments (incl. strategy="audioset", extra ml), text_before_segments, beat_grid/BeatGrid (extra beats)
mixing.video	Video, crop_video, loop_video, change_speed, replace_audio, normalize_audio, overlay_ambient_bed, ken_burns_video/ken_burns_film, concatenate_videos (+ transitions: crossfade_transition, fade_through_black, …), make_thumbnail, write_subtitles_in_video, SOCIAL_SIZES
mixing.video.genai	generate_video (Vertex AI Veo)
mixing.transcript	transcribe, remove_fillers, srt_for_media, words_to_srt/words_to_prose
mixing.chapters	detect_chapters, Chapter
mixing.dubbing	text_to_speech, list_voices, translate_srt, dub_video_from_srt
mixing.srt	Cue, parse_srt, dump_srt, seconds_to_srt_time

See the skills for the full, example-driven API of each.

License

MIT