Google Vids AI voiceover: the same workflow through an API

Google Vids is a point-and-click editor. If you want the same voiceover step in code, write the script, generate speech with a Sume TTS job, and lay the audio under your clips in Timeline 1.0. That gives you a repeatable pipeline with job ids, not a session in an editor.

Releasebot's Google page (read 2026-10-02) lists the 2026-10-02 Workspace weekly recap, which mentions more natural voiceovers from an upgraded Gemini 3.8 Flash Lite TTS in Google Vids, and customizable captions. Releasebot is a third-party aggregator, so read this as reported, not as Google's own wording. Sume's TTS is a different product with its own voices; see the engine note below.

What are the steps?

Four calls cover it. The table lists each, with the Sume surface.

Which engine does Sume's TTS use?

Sume's TTS 1.0 has no engine picker: model and model_id are rejected with 400. The separate TTS Router lists explicit models at GET /v1/tts-router/models. Check that catalog for what is routable today rather than assuming it matches an engine named in Google's recap.

What output format should I ask for?

The default is mp3 at 44100 Hz and 128 kbps. When the audio will feed an avatar mux, the OpenAPI notes say to pass wav with pcm_s16le at 44100 explicitly. For a plain video voiceover the default is fine.

Can Sume match Vids captions?

Sume has its own caption styles through POST /v1/video-captions; see the companion post on aligning captions to a script. It does not copy Vids' caption options.

Where does a scripted pipeline beat an editor?

An editor is quick for one video. A pipeline pays off at the tenth: the script is a file, the voice is a parameter, and the render is a job with an id. When a price or a claim changes, you edit one line and re-run, instead of reopening a project and re-recording.

It also helps review. A script in version control has a history, and each voiceover job records the transcript it was generated from. A reviewer reads the text, and the audio follows from it.

What does a minimal run look like?

Generate each paragraph as its own TTS job, keep the ids in order, join the audio with a concat job, and send the joined file to Timeline as the audio spine. If a paragraph reads badly, regenerate only that one and join again. The sentence segmentation option can return gapless per-sentence slices when you need to line visuals up with sentences.