Choose a text-to-speech model by the controls your script needs: emotion and pronunciation, dialogue between speakers, continuity across long narration, or the lowest price per character. The models below share one billing unit on Raywake, characters of input text, but expose very different request fields.
This comparison uses the request contracts available on Raywake on 23 October 2026. It compares documented controls and quoted prices, not voice quality or latency, which depend on your script, language and listener.
Which controls does each model expose?
| Model | Text limit in the contract | Voice | Distinctive controls |
|---|---|---|---|
| ElevenLabs v4 Turbo | 5,000 characters | Named voice (default Rachel) | Stability, similarity boost, ISO 639-1 language code, timestamps, 19 output formats |
| ElevenLabs v4 | 5,000 characters | Named voice (default Rachel) | Same controls as v4 Turbo |
| ElevenLabs Multilingual v2 | Not stated in the contract | Named voice (default Rachel) | Previous and next text for continuity, speed 0.7–1.2, style, stability, timestamps |
| MiniMax Speech-02 HD | 5,000 characters | Voice ID (default Wise_Woman) | Seven emotions, pitch −12 to 12, speed 0.5–2, pronunciation dictionary, language boost |
| MiniMax Speech 2.8 HD | 10,000 characters | Voice ID | Pause tags, interjections such as (laughs), timbre and intensity, loudness normalization |
| Gemini 3.8 Flash TTS | 8,192 tokens for the whole input | 30 prebuilt voices | Multi-speaker turns, style instructions separate from the transcript |
A control that is missing from a contract cannot be set through that request; it is not a statement about what the underlying model could do elsewhere. Sources: Eleven v4 Turbo, Eleven v4, Multilingual v2, Speech-02 HD, Speech 2.8 HD, Gemini 3.8 Flash TTS.
What does 1,000 characters cost?
| Model | Raywake quote for 1,000 characters |
|---|---|
| ElevenLabs v4 Turbo | up to $0.07 |
| ElevenLabs v4 | up to $0.14 |
| Gemini 3.8 Flash TTS | up to $0.16 |
| ElevenLabs Multilingual v2 | up to $0.18 |
| MiniMax Speech-02 HD | up to $0.18 |
| MiniMax Speech 2.8 HD | up to $0.18 |
These are the quote ceilings for a 1,000-character request on 23 October 2026. A quote freezes the price before you generate, and the final charge can settle below it when the provider bills less. A 4,000-character script is roughly four times the figure above. Request a quote with your actual text, because characters, not words or seconds, are the billing unit.
Choose by script requirement
A product video or app voice-over in English. Start with ElevenLabs v4 Turbo. It is the least expensive model here, accepts 5,000 characters per request and can return timestamps alongside the audio when you need captions. Move to v4 when you want to compare a second rendering of the same voice and settings.
Long narration assembled from several requests. Multilingual v2 accepts the text that comes before and after the current request. Passing neighbouring sentences helps a chapter split into parts sound continuous instead of restarting its intonation at every boundary.
Emotion, numbers and names that must be read correctly. MiniMax Speech-02 HD exposes an emotion choice, a pronunciation dictionary and a language boost that includes Turkish, Arabic, German, Japanese and Hindi among others. Use the dictionary for product names and use English normalization when a script contains figures.
A script with pauses and reactions. MiniMax Speech 2.8 HD reads pause tags such as <#0.5#> and interjections such as (laughs) or (sighs) inside the text, accepts 10,000 characters and can normalize loudness for a consistent level across files.
A conversation between two people. Gemini 3.8 Flash TTS accepts a list of speakers, each with its own voice, and an ordered list of turns. Delivery directions such as "calm and friendly" go in style instructions, separate from the words that are spoken.
How to compare them fairly
Use the same script, language and target output format for every model. Keep a short sample that contains names, numbers and a question, because those expose pronunciation and intonation differences quickly. Listen on the device your audience uses before choosing a voice for a whole project.
Save the voice, settings and seed you approve. Changing stability, speed or emotion between files makes a series sound inconsistent even when the model stays the same.
If the audio will be paired with video, generate the approved voice track first and use it in a lip-sync or video model, rather than relying on a video model to write and speak exact dialogue. The native-audio video guide explains that distinction.
Try a text-to-speech model