HyperFrames Media Preprocessing

Three CLI commands that produce assets for compositions: tts (speech), transcribe (timestamps), and remove-background (transparent video). Each downloads a model on first run and caches it under ~/.cache/hyperframes/. Drop the output into the project, then reference it from the composition HTML — see the hyperframes skill for the audio/video element conventions.

Text-to-Speech (tts)

Generate speech audio locally with Kokoro-82M. No API key.

npx hyperframes tts "Text here" --voice af_nova --output narration.wav
npx hyperframes tts script.txt --voice bf_emma --output narration.wav
npx hyperframes tts --list                       # all 54 voices

Voice Selection

Match voice to content. Default is af_heart.

Content type	Voice	Why
Product demo	af_heart/af_nova	Warm, professional
Tutorial / how-to	am_adam/bf_emma	Neutral, easy to follow
Marketing / promo	af_sky/am_michael	Energetic or authoritative
Documentation	bf_emma/bm_george	Clear British English, formal
Casual / social	af_heart/af_sky	Approachable, natural

Multilingual

Voice IDs encode language in the first letter: a=American English, b=British English, e=Spanish, f=French, h=Hindi, i=Italian, j=Japanese, p=Brazilian Portuguese, z=Mandarin. The CLI auto-detects the phonemizer locale from the prefix — no --lang needed when the voice matches the text.

npx hyperframes tts "La reunión empieza a las nueve" --voice ef_dora --output es.wav
npx hyperframes tts "今日はいい天気ですね" --voice jf_alpha --output ja.wav

Use --lang only to override auto-detection (stylized accents). Valid codes: en-us, en-gb, es, fr-fr, hi, it, pt-br, ja, zh. Non-English phonemization requires espeak-ng system-wide (brew install espeak-ng / apt-get install espeak-ng).

Speed

• 0.7-0.8 — tutorial, complex content, accessibility
• 1.0 — natural pace (default)
• 1.1-1.2 — intros, transitions, upbeat content
• 1.5+ — rarely appropriate; test carefully

Long Scripts

For more than a few paragraphs, write to a .txt file and pass the path. Inputs over ~5 minutes of speech may benefit from splitting into segments.

Requirements

Python 3.8+ with kokoro-onnx and soundfile (pip install kokoro-onnx soundfile). Model downloads on first use (~311 MB + ~27 MB voices, cached in ~/.cache/hyperframes/tts/).

Transcription (transcribe)

Produce a normalized transcript.json with word-level timestamps.

npx hyperframes transcribe audio.mp3
npx hyperframes transcribe video.mp4 --model small --language es
npx hyperframes transcribe subtitles.srt          # import existing
npx hyperframes transcribe subtitles.vtt
npx hyperframes transcribe openai-response.json

Language Rule (Non-Negotiable)

Never use .en models unless the user explicitly states the audio is English. .en models (small.en, medium.en) translate non-English audio into English instead of transcribing it. This silently destroys the original language.

1. Language known and non-English → --model small --language <code> (no .en suffix)
2. Language known and English → --model small.en
3. Language unknown → --model small (no .en, no --language) — whisper auto-detects

Default model is small, not small.en.

Model Sizes

Model	Size	Speed	When to use
tiny	75 MB	Fastest	Quick previews, testing pipeline
base	142 MB	Fast	Short clips, clear audio
small	466 MB	Moderate	Default — most content
medium	1.5 GB	Slow	Important content, noisy audio, music
large-v3	3.1 GB	Slowest	Production quality

Music with vocals: start at medium minimum; produced tracks often need manual SRT/VTT import. For caption-quality checks (mandatory after every transcription), the cleaning JS, retry rules, and the OpenAI/Groq API import path, see hyperframes/references/transcript-guide.md.

Output Shape

Compositions consume a flat array of word objects. The id field (w0, w1, ...) is added during normalization for stable references in caption overrides; it's optional for backwards compatibility.

[
  { "id": "w0", "text": "Hello", "start": 0.0, "end": 0.5 },
  { "id": "w1", "text": "world.", "start": 0.6, "end": 1.2 }
]

Background Removal (remove-background)

Remove the background from a video or image so the subject (typically a person — avatar, presenter, talking head) sits as a transparent overlay in a composition.

npx hyperframes remove-background subject.mp4 -o transparent.webm  # default: VP9 alpha WebM
npx hyperframes remove-background subject.mp4 -o transparent.mov   # ProRes 4444 (editing)
npx hyperframes remove-background portrait.jpg -o cutout.png       # single-image cutout
npx hyperframes remove-background subject.mp4 -o subject.webm \
  --background-output plate.webm                                   # both layers in one pass
npx hyperframes remove-background subject.mp4 -o transparent.webm --device cpu
npx hyperframes remove-background --info                           # detected providers

Uses u2net_human_seg (MIT). First run downloads ~168 MB of weights to ~/.cache/hyperframes/background-removal/models/.

Do NOT use u2net for die-cut stickers with a white outline

remove-background (u2net) is salient-object segmentation — it segments the subject body and cuts off a white die-cut outline, because the white stroke around the figure is not "the object". For transparent Telegram / messaging stickers rendered with a white outline on a flat known-colour background, key the background instead of segmenting the subject:

• Flood-fill the flat bg. PIL ImageDraw.floodfill seeded from points along all four edges (step ~24px), mark filled pixels → alpha 0, everything else opaque. This preserves the white outline AND interior same-colour props AND disconnected islands (a flying cup separate from the body keeps its own outline). Requires the gen prompt to put the subject on a flat uniform bg and ask for the white die-cut outline explicitly.
• No-outline variant (cleaner than stripping a baked outline): re-run each image image-to-image — pass the finished render as --ref and prompt "recreate identical pose/props but REMOVE the white outline and put it on a flat pure chroma-key green (#00b140)", then key the green with a soft chroma key, never a binary mask (alpha = clip((HIGH-(G-max(R,B)))/(HIGH-LOW),0,1), LOW≈30 HIGH≈95, plus global despill G = min(G, max(R,B))). A binary flood-fill + erosion produces staircase / aliased edges — blur the binary alpha (GaussianBlur(0.8)) for a smooth edge instead of eroding.
• Sizing for messaging stickers: crop to the solid alpha bbox and scale the long side to exactly the target (e.g. 512) — do not pad to a 512² square (it leaves the subject small with empty space around it).

Does NOT apply to: human cutouts for video overlays (u2net is correct there); art without a die-cut outline; subjects on a non-uniform / busy background (flood-fill needs a flat key colour).

Layer separation (--background-output)

Pass --background-output (or -b) to emit a second transparent video alongside the cutout: same source RGB, alpha is 255 − mask instead of mask. The cutout is the subject with a transparent background; the plate is the original surroundings with a transparent hole where the subject was.

File	Alpha is…	Use it for
-o subject.webm	The mask — subject opaque, background transparent	Foreground layer, place on top
--background-output plate.webm	Inverse — surroundings opaque, subject region transparent	Bottom layer; put text or graphics between this and the subject

Both outputs share the same --quality preset and run from a single inference pass — encode cost roughly doubles, segmentation cost stays the same. Only valid for video inputs and .webm/.mov outputs.

Hole-cut plate, not an inpainted clean plate. The subject region in plate.webm is fully transparent — composite something opaque under it to fill the hole. The single test for whether --background-output is the right tool: will anything ever be visible through the subject's silhouette where the subject used to be?

Use case	Right tool
Text/graphics between the cutout and the plate (this command's reason for existing)	Hole-cut (--background-output)
Subject onto an unrelated scene	Just subject.webm; ignore the plate
Show the room without the person, alone over no other content	Clean plate — needs an inpainter (LaMa, ProPainter, E2FGVI). Not this command.
Replace the subject with a different subject	Clean plate — same as above

If a user asks for "the room with the person removed" and intends to display it standalone, do not reach for --background-output. Tell them they need an inpainter.

Typical layered composition (the canonical hole-cut use case):

<!-- z=1 the inverse-alpha plate fills everything except the subject region -->
<video
  src="plate.webm"
  data-start="0"
  data-duration="6"
  data-track-index="0"
  muted
  playsinline
></video>

<!-- z=2 graphics / text live between the two layers -->
<h1 id="headline" style="z-index:2; ...">MAKE IT IN HYPERFRAMES</h1>

<!-- z=3 the cutout floats the subject back over the headline -->
<div class="cutout-wrap" style="position:absolute;inset:0;z-index:3">
  <video
    src="subject.webm"
    data-start="0"
    data-duration="6"
    data-track-index="1"
    muted
    playsinline
  ></video>
</div>

This is functionally equivalent to the text-behind-subject pattern below, but you don't need the original presenter.mp4 in the project — the plate replaces it. Useful when you want to ship just the two transparent layers and let the user drop arbitrary content between them.

Output Format

Format	When
.webm (VP9 + alpha)	Default. Compositions play this directly via <video>.
.mov (ProRes 4444)	Editing in DaVinci/Premiere/FCP. Large files.
.png	Single-image cutout (still subject, layered over a backdrop).

Chrome decodes VP9 alpha natively, so the .webm plugs into a composition like any other muted-autoplay video — see the hyperframes skill for the <video> track conventions.

Quality presets

--quality fast|balanced|best controls only the VP9 encoder's CRF — segmentation quality is fixed.

Preset	CRF	When
fast	30	Iterating, smaller file, looser color match
balanced	18	Default. Visually identical for most uses
best	12	Master / final delivery. Largest file, tightest match

Compositing patterns — pick the right one

The cutout webm is a re-encoded copy of the source mp4's RGB. That choice has consequences depending on what you put behind it:

Pattern	What's behind the cutout	Result
Cutout over a different scene (most common)	Static image, gradient, or unrelated video	Looks great. The cutout's RGB is the only source of the subject — no doubling, no edge halo. This is what remove-background is built for.
Cutout over its own source mp4 (text-behind-subject)	Same mp4 the cutout was generated from	Two RGB sources for the same person. At default --quality balanced (crf 18) the doubling is barely visible; at --quality fast (crf 30) you'll see a faint color shift / edge halo. Use --quality best (crf 12) for masters.
Cutout over a different take of the same person	Footage of the same subject	Will look like two separate people overlapping. Don't do this.

Text-behind-subject (headline behind a presenter):

<video
  src="presenter.mp4"
  id="bg"
  data-start="0"
  data-duration="6"
  data-track-index="0"
  muted
  playsinline
></video>
<h1 id="headline" style="z-index:2; ...">MAKE IT IN HYPERFRAMES</h1>
<div class="cutout-wrap" style="position:absolute;inset:0;z-index:3;opacity:0">
  <video
    src="presenter.webm"
    data-start="0"
    data-duration="6"
    data-track-index="1"
    muted
    playsinline
  ></video>
</div>

Two key rules:

1. Wrap the cutout video in a non-timed <div> and animate the wrapper's opacity, not the video element's. The framework forces opacity:1 on active clips (any element with data-start/data-duration), so animating the video's opacity directly is silently overridden. The wrapper has no data-* attributes, so it's owned by your CSS/GSAP.
2. Both videos use data-start="0" and data-media-start="0" so the framework decodes them in sync from t=0. Late-mounting the cutout (data-start=3.3) introduces a seek + warm-up that lands a frame off the base mp4 — visible as one frame of misalignment at the cut.

Then GSAP-flip the wrapper opacity at the cut: tl.set(cutoutWrap, { opacity: 1 }, 3.3).

TTS → Transcribe → Captions

When there's no pre-recorded voiceover, generate one and transcribe it back to get word-level timestamps for captions:

npx hyperframes tts script.txt --voice af_heart --output narration.wav
npx hyperframes transcribe narration.wav   # → transcript.json

Whisper extracts precise word boundaries from the generated audio, so caption timing matches delivery without hand-tuning.