ElevenLabs: Eleven Multilingual v2

Eleven Multilingual v2 is ElevenLabs’ text-to-speech model for natural, emotionally aware and consistent narration. It accepts a written script and returns spoken audio, with the selected voice supplying much of its character and accent. The producer highlights stability on longer generations, professional content, character voiceovers, e-learning and multilingual projects. It is a speech-synthesis model, separate from the similarly named Multilingual Speech to Speech voice changer.

The documented set contains 29 languages: English, Japanese, Chinese, German, Hindi, French, Korean, Portuguese, Italian, Spanish, Indonesian, Dutch, Turkish, Filipino, Polish, Swedish, Bulgarian, Romanian, Arabic, Czech, Greek, Finnish, Croatian, Malay, Slovak, Danish, Tamil, Ukrainian and Russian. Prepare the actual script in a supported language before synthesis rather than expecting the voice model to translate it.

The model is designed to retain the chosen speaker’s characteristics and personality through supported languages. Accent still depends strongly on the voice: ElevenLabs recommends a voice trained in the target language and regional accent, because a non-native voice may retain its original accent or drift. For a localised course, preserve the narrator’s role and overall tone while adapting sentence structure, terminology and pronunciation to the audience. A shared voice identity does not remove the need to review the target-language delivery.

The request limit is 10,000 characters. Longer material can be divided into connected chapters or sections; the producer’s shared TTS workflow documents previous and next text, or neighbouring request context, for preserving prosody between chunks. Keep the division at a meaningful sentence boundary and carry enough surrounding context for the start and end of a passage. The manufacturer’s approximate minutes per request are not a fixed-duration promise, because words, language and pace determine the selected recording’s length.

Delivery follows the text’s meaning, narrative context, punctuation and the selected voice. Stability controls variability and emotional range; similarity controls how closely the result resembles the voice. The shared product guide describes a speed range of 0.7 to 1.2 with 1.0 unchanged. High stability can sound restrained or monotonous, while too little stability can make a performance erratic. Written delivery prose may itself be spoken, so prepare the final script rather than embedding instructions that should never be heard.

This model’s pause workflow supports SSML break tags, unlike v3 and v4. The producer gives <break time="1.5s" /> as a way to request a natural pause and documents pauses up to three seconds. Excessive break tags can introduce speed changes or artifacts, and the selected voice can affect how a pause sounds. Use breaks at meaningful transitions, such as after a definition or before the next part of a narrated instruction.

Multilingual v2 does not support phoneme tags in the manufacturer’s pronunciation guide. Use dictionary aliases or a prepared alternative spelling to indicate the desired spoken form of a name, acronym or technical term. ElevenLabs also describes this model as handling numbers more naturally than Flash v2.5, while its preparation guidance still recommends writing critical numbers, dates and symbols as the words that should be spoken. A reference code and a quantity need different readings even when their written digits look similar.

The producer audio format family includes MP3 for distribution, uncompressed 16-bit PCM for production work, μ-law and A-law for telephony and Opus. Full speech generation and streaming are documented in the shared TTS workflow. Choose the output for the intended destination and retain the script, voice and selected recording together. An audiobook chapter, a corporate narration and a short lesson can use the same model purpose while requiring different pacing and segmentation.

For a practical lesson, write the introduction, definitions, example and transition as complete spoken passages. Mark the terms that need consistent pronunciation and use a natural pause after the core concept. Review a representative part that contains names, numbers and a longer sentence, checking accent, intelligibility and whether the delivery remains steady through the section boundary. The useful artifact is a consistent narration recording that can be placed into the course, audiobook or video alongside its source script.