Voice models
The Voiceover node (Озвучка) turns exact text into an audio asset using a selected voice. It is separate from the native speech a video model might create. A voice identifier chooses the speaker; a model chooses the synthesis model. Connecting a voice or text input overrides the corresponding local field.
Multilingual v2
Section titled “Multilingual v2”The default Voiceover choice. Use it for a straightforward spoken script; still listen for pronunciation and omitted words.
| Setting | Value |
|---|---|
| Name | Multilingual v2 (дословно, RU) |
| Availability | available |
| Default for new nodes | yes |
Eleven v3
Section titled “Eleven v3”An expressive speech option. Performance can vary, so review emotional delivery and the literal words together.
| Setting | Value |
|---|---|
| Name | v3 (выразительно) |
| Availability | available |
| Default for new nodes | no |
Flash v2.5
Section titled “Flash v2.5”A faster-oriented speech option. Spell out ambiguous dates, numbers and abbreviations when their spoken form matters.
| Setting | Value |
|---|---|
| Name | Flash v2.5 (быстро) |
| Availability | available |
| Default for new nodes | no |
Studio example
Section titled “Studio example”Connect the intended Voice and the text to Voiceover. Enter: “The greenhouse opens at eight in the morning.” Select a model and review the voice settings, then generate the audio. Listen to the whole sentence, confirm the speaker and language, and measure whether it fits the intended scene before assembly.
Write only the words the audience should hear in the script. Keep visual direction and production notes outside that text. A text line is not a timeline: a long line can exceed the video slot even when the synthesis request succeeds.
Failure modes and evidence
Section titled “Failure modes and evidence”Check names, numbers, stress, pauses and clipping. A convincing voice can still say the wrong word. A multilingual model name does not promise equal pronunciation across every voice and language. This page verifies the three currently selectable records and the input precedence; no new listening benchmark was run for this guide.
Other ElevenLabs models in the author’s catalog are not implied to be connected. Voice cloning, dubbing and isolation are separate Studio operations rather than additional choices in this TTS selector. See narration, dialogue and the MCP support table for their execution boundary.
Reviewed 14 October 2026 against the Studio selector and ElevenLabs model documentation. Recheck the script and settings when changing model or voice.