Eleven v4 stacked tags: direct emotion without them

TechCrunch reported on September 28, 2026 that ElevenLabs v4 can stack inline expression tags and clone a voice from 10 seconds of audio. Sume's docs list a tts_create tool but no voice-cloning route, so on Sume you direct delivery through the script itself and check the tool's contract first.

What v4 changed, per the report

• Clones a voice from 10 seconds of audio.
• Languages up from 70 to 90.
• Improved Japanese, Brazilian Portuguese, Mandarin and Cantonese.
• Inline tags can be stacked, and the model follows the sequence.

What Sume documents

The hosted MCP lists tts_create among paid generation tools, and tts_source_get and tts_source_verify_spine as free reads for accepted-script checks. The docs I read do not describe tag syntax, cloning or a voice list, so do not assume a v4-style tag works. Call tools_schema with name: "tts_create" to read the live contract before sending a script.

Call tools_schema with name "tts_create" and list every field
that controls voice, language and script before submitting.

Direct a voice without tags

Delivery is mostly decided by the text. These habits work with any speech model and need no special syntax:

• Write short sentences; one idea each.
• Use punctuation for pacing: commas for breath, full stops for finality.
• Spell out numbers and abbreviations the way you want them said.
• Render a line at a time, then keep the takes you like; a wave of parallel creates is what script_run and batch jobs_wait are for.
• Listen, then rewrite the script instead of fighting the output.

Talking faces

If the voice drives an on-camera person, the Sume docs say a talking face is a Fabric clip made from an accepted still plus the TTS audio. Video models do not lip-sync to narration added later.