One upload.
Every output.

Five models. One workflow.

Sliver isn't one AI doing tricks — it's a pipeline of specialized models, each pointed at the part of the cut they're actually best at. The result feels like a human editor because every stage is doing the human job in miniature.

• transcribe

  Whisper-large-v3

  Speech → text + word-level timestamps at ~20× realtime on an A10 GPU. The foundation everything else builds on.

• diarize

  WhisperX + pyannote

  Audio diarization. Tells us which host is speaking when, so the reframe + caption color follow the right person across a multi-host frame.

• score

  Claude Haiku 4.5

  Reads the transcript by 10-minute windows, picks the moments worth clipping, tiers their target length to the structure (15-30s reaction, 60-90s setup-payoff).

• see

  Claude vision

  Samples frames every ~5 seconds for on-screen events the transcript misses — kill streaks, screen-share switches, reaction shots, P&L flips.

• refine

  Claude Haiku 4.5 (tool-use)

  The natural-language refinement layer in the timeline editor — "extend this 5 seconds," "punchier captions," "stop tracking the speaker." Outputs a JSON patch, the renderer re-cuts just the changed piece.

Your source is processed once, stored encrypted, never used to train models, and deleted after 30 days — Business plans can request immediate purge.

Why Sliver

Built for podcast + multi-person frames

Speaker-tracked crops.
Speaker-colored captions.

Center-crop fails the moment two people share a frame. Sliver runs WhisperX diarization on the audio, cross-references with face detection on the video, and reframes the 9:16 short around whoever is actually talking. Caption color follows the speaker too — viewers track who said what without you touching a single style.

Who it’s for

Use cases.

Sliver isn’t a niche-vertical tool. It’s built for any creator turning long-form recordings into something publishable — with the speaker-tracking + AI refinement that competitors save for their next quarter.

1. 01

   Long-form podcasters

   Drop in a 90-minute episode, get back chapter splits, a condensed cut, and 8-15 speaker-tracked shorts ready for TikTok and Reels — without scrubbing the timeline yourself.

2. 02

   Live streamers + gaming creators

   Upload an 8-hour Twitch VOD. Vision-scoring catches the actual reaction moments (kill streaks, P&L flips, face-cam laughs) instead of mid-monologue lulls. Static-crop competitors miss these.

3. 03

   Interview + talk-show hosts

   Multi-person frames are where every other tool falls apart. Sliver reframes the 9:16 short around whoever is talking and colors captions by speaker so viewers always know who said what.

4. 04

   Course creators + educators

   Turn long lectures into chapter-split YouTube episodes plus 60-90s “key concept” shorts. The natural-language refinement panel rewrites a cut in one prompt — “tighten the explanation” or “cut the tangent”.

5. 05

   Agencies + repurposing teams

   One creator’s stream becomes a week of content across four platforms. Bulk-publish to YouTube Shorts, TikTok, Reels, and YouTube long-form from a single project. API access available on the Business plan.

6. 06

   Founders + execs doing content

   Don’t have time to edit? Drop a recording, get shareable clips back. AI rewrites titles for each platform’s native style so you don’t reword the same hook four times.

Plans

Start free.

Two videos a month on Free, no card required. Move to Starter ($19), Pro ($39), or Business ($99) when your volume picks up — 7-day trial on every paid tier.