Annotate the script with speaker names. The Agent maps each speaker to a voice through `vb tts`, renders the lines, drives ffmpeg to lay them on a timeline with BGM that ducks under speech, and exports a master MP3 plus per-speaker stems. Apple-Podcasts-grade loudness by default.
Picked from `vb tts providers`. Saved with this episode — character recurs in episode 13, same voice. No drift, no re-explaining.
Anchors the conversation · neutral US English · light pacing
Brings the heat · expressive · code-switches into Mandarin cleanly
Sidechain ducks under speech · loops cleanly to fill 28 min
Speech segments push the BGM down 8 dB underneath them — drawn here for the first 35 seconds so you can see exactly how the mix is built.
Tag who says what — "[Host]:", "[Guest]:", or your own naming. Plain text, Markdown, .docx, even Fountain screenplay format all work. The Agent infers tone, age, energy from the surrounding text and picks voices accordingly.
Two-host podcast with BGM? 30s commercial in three lengths? Audiobook chapter with separate character voices? Tell the Agent what the deliverable is and the right pipeline (BGM ducking / loudness target / chapter markers) snaps into place.
A zip lands in your Drive: master.mp3 ready to ship, per-speaker stems for the editor, a separate BGM stem, and a loudness compliance report (integrated, true peak, range). Hand the stems to your editor for tweaks — you do not need to re-run from scratch.
Most AI voiceover tools concatenate clips with ugly transitions. The Agent runs ffmpeg with sidechain compression so the BGM ducks under speech, applies fade in/out on every line, and equalizes loudness across speakers so the guest doesn't suddenly sound twice as loud as the host.
The Agent reads the script context ("the founder is in her 40s, calm but direct") and picks a voice that fits — and once picked, the same character keeps that voice across episodes. The mapping is saved as a small `.json` so the next chapter / next episode is automatically consistent.
Each delivery target has a different loudness spec — Apple Podcasts wants -16 LUFS integrated, broadcast wants -23, social videos want -14. Pick the target and the Agent measures the final mix and re-renders until it passes. The compliance report is included in the zip so the platform never bounces your upload.
You can take the host stem, the guest stem, and the BGM stem straight into ProTools / Logic / Audition. No "AI black box" — every layer is exposed, every level is documented. The Agent did the boring 80%; your editor adds the last 20% of polish.
Whichever is best for the language and tone. `vb tts` exposes ElevenLabs (warm, English-dominant), OpenAI tts-1-hd (clean, multilingual), 字节豆包 (best Mandarin / Cantonese / dialects), and others — and the Agent picks based on the script. You can pin a provider in chat ("use ElevenLabs only") and `vb tts providers` lists everything available.
Sign in, hand it over to the Agent — the finished file lands in your Drive.