AI Podcast Production: Record, Edit, Mix and Export in One Session

Podcast post-production is one of the most repetitive audio editing tasks that exists. Every episode involves the same operations: noise removal, silence trimming, level normalization, music bed mixing, chapter markers, and export. AI audio editors can automate each of these without requiring you to learn a DAW.

The Traditional Podcast Workflow (and Why It Takes So Long)

A typical solo-host podcast episode runs 30–60 minutes. Editing a raw recording to a publishable episode in a traditional DAW involves:

• Manual review of the waveform to find and cut long pauses.
• Noise gate or spectral repair to remove room noise and HVAC hum.
• Loudness normalization to a delivery target (-14 LUFS for Spotify and YouTube, -16 LUFS for Apple Podcasts, -23 LUFS for broadcast).
• Music intro/outro mixing with level automation.
• Export to MP3 at appropriate bitrate with ID3 tags.
• Show notes generation from timestamps.

Each of these steps requires different tools, different knowledge, and careful listening. Together they make editing an episode a long job, and most of it is the same moves every time.

The AI-Augmented Workflow

An AI audio editor with natural language control can replace most of this with a single session. Here is a realistic workflow using edytlab:

Step 1: Load the Raw Recording

Drag your WAV file into the session or type "load episode-045-raw.wav". The agent adds it as the first track. If you have a separate music bed file, load that too.

Step 2: Transcribe and Review

From v0.8.0 the step is one sentence, "transcribe track 1". The first time, you download the speech model (Whisper base, about 139 MB) in Settings → Account; nothing downloads on its own. Then the agent runs Whisper on your computer, with no upload of audio and no API key for transcription, and adds a word-level transcript with timestamps to the session, so you can see where filler words, long silences and retakes are without scrubbing the waveform. The words come back to the AI provider you chose as text, a bounded amount for each call, so the assistant can work with them; with Ollama they stay on your computer. Ducking the music under the speech in step 4 does not need a transcript: it keys on the voice track's audio. On v0.7.0 and earlier the transcribe tool returns an error, so find retakes by listening.

Whisper runs entirely on your computer, on the CPU, and the word-level transcript is stored in the session. How long it takes depends on your processor: on the machine we measured, seven seconds of audio took about three seconds the first time, and the tool's own description says usually under half the audio's length. The whole recording is decoded into memory first, so an hour of stereo audio at 44.1 kHz takes about 1.3 GB of it, and we have not timed an episode that long. The model's words are its best guess: read them before you cut on them.

Step 3: Describe the Edits

Describe what you want, in plain English: "Cut all silences longer than 1.5 seconds. Remove the section between 12:30 and 13:45 — that was an off-topic tangent. Normalize to -16 LUFS." The agent executes each operation as a tool call against the session DAG.

Step 4: Mix Music Beds

Load your intro/outro music: "Add intro.wav to track 2, crossfade into the speech at 0:08, and duck the music under the speech to -18 dB". The agent handles the volume automation and crossfade geometry. You can preview immediately.

Step 5: Export

Type "export as MP3 192kbps with title Episode 45, author My Podcast". Done. The session state is saved as a DAG, so you can branch it, revert any edit, or export different versions (clean edit vs. explicit version) without re-doing work.

What AI Cannot Replace (Yet)

Automated workflows do not replace critical listening. AI can normalize to a target LUFS, but it does not know if your interview guest had an unusually nasally recording environment that day. From v0.8.0 the assistant can find the ums and filler words in a transcript and remove them (it reports what it found first and changes nothing until you say so), but rhythm editing — making the conversation flow more naturally — still benefits from a human ear. Use AI to handle the mechanical steps and spend your time on the creative ones.

Multi-Guest Podcast Editing

For interviews with multiple speakers, load each recording as a separate track. When you only have a mixed recording, split_by_speaker splits it into one track per speaker from speaker segments you give it, so each voice can be normalized and treated on its own, then re-mixed. This is not a perfect substitute for separate track recording, but it is production-viable for remote interviews recorded on a single channel.

edytlab is an open-source, local-first AI audio editor. Download the latest release or star it on GitHub.