Zero Shot

The same woman, sharp once the fog clears

FROM A RAW TAPE TO ONE MASTER

Listen to the demo

Queue

Try it out yourself

Status off

Add a YouTube URL and magically turn the audio from it into a mastered podcast episode. It works best on interview or topic explainer videos.

Status is off. No new items will be accepted.

The queue

Each row is one accepted video. The first column is the URL that was submitted. The second column is the video title, filled in after lookup.

No URLs yet.

Accepted videos. The first column is the URL. The second column is the video title.
URLTitle

Apple Podcasts

Zero Shot Podcast

When a podcast episode is created, it automatically uploads to this Apple Podcast. It generally takes a few hours for an episode to be processed and uploaded.

Listen on Apple Podcasts

One messy tape. One master.

A long interview comes in overlapping, noisy, and full of asides. The app finds the ideal way to say what was actually spoken, keeps the golden stretches, and weaves them with grounded voice and music into master.wav.

Custom pipelines

We can make custom pipelines for any creator, based on their needs. This demo is free for now, but it showcases how the app works.

pipeline.run

From the source to the master

Five passes take the source audio and hand back a finished episode.

  1. HearWe transcribe all of the source.
  2. ReadWe find which segments to keep and which to discard.
  3. ShapeWe shape it into a story, with an optional synthetic interviewer voice to weave it together.
  4. BuildWe add the music.
  5. ShipWe push the final master.
Hear

Zero Shot Media takes a raw interview and builds the master. The story is chosen for this tape. The audio work stays on the machine.

A dark mixing desk beside the mission

Language models decide the shape and write only from what was said. Local models hear the tape, speak the new lines, and write the music. Speech stays above the bed.

See the method

How analysis works

After the words are on the page, the run reads the tape for what a first listener should keep.

See the run

How it looks under the hood

The packages under the five passes. Speech, speakers, voice, and music run on this machine.

  • Whisper

    mlx-audio

    ffmpeg turns the upload into 16-bit WAV. mlx-community/whisper-large-v3-turbo then writes every word, with timings. DeepFilterNet3 can quiet the room first. If that stack is unavailable, ffmpeg afftdn is the fallback.

  • Sortformer

    Four speakers

    mlx-community/diar_sortformer_4spk-v1-fp32 marks who is talking. The model covers up to four voices, and it runs on MLX beside Whisper.

  • OpenAI

    gpt-5.6-terra

    gpt-5.6-terra is the flagship for the reading, the shape, and the bridge lines. Smaller calls use gpt-4o-mini. gpt-4o conducts the walk. The cover is three gpt-image-2 candidates.

  • Chatterbox

    ResembleAI

    ResembleAI/chatterbox speaks the optional interviewer lines in a voice already on the tape. If synthesis fails, mlx-audio is the fallback voice.

  • MusicGen

    Large, then down

    Theme beds start at facebook/musicgen-large, then musicgen-medium, then musicgen-small. Melody beds use musicgen-melody. MMAudio is the backup. The mix keeps the talk above the bed in master.wav.

About

Nicket Uttarwar is an artist and a developer. He built Zero Shot Media for the custom editing pipeline he needed himself, and he wants to share it.

nicketuttarwar.com

Your show

A pipeline built for one creator

This is for people outside the media industry, and for individual creators who have a lot to say and no time to learn audio engineering. You upload the recordings. The pipeline is shaped to your voice and your show, and it returns a finished master.

It is a free service for now. Write nicket@uttarwar.com.