Dia when you want a short exchange and nothing to configure
This engine has no voice field. There is no list to choose from and there are no sliders — you write the copy with labelled turns and get a file with two people in it. Zero configuration is at once its whole advantage and its whole drawback: you cannot ask for a particular timbre. It will not read Polish.
Rendered by this model
Turn the sound on and listen
The script the engine was given
[S1] The new collection is in the shop today. [S2] All of it? (laughs) [S1] All of it. Cotton, linen, and the colour that sold out last time. [S2] Then order before two, and it ships the same day.
Specification
- Built by
- Nari Labs
- Speaks Polish
- no — reads Polish text phonetically
- Ready-made voices
- none — speakers are set with tags in the script
- Voices in one file
- up to 2
- Steering the delivery
- tags written into the script
- File from the engine
- wav — re-encoded to mp3 for this page
- Render time
- about 120s
- Cost
- 1 credit per 1000 characters, rounded up
What it is good for
- There is nothing to pick
- An empty job apart from the text itself. Speaker tags are the whole of the control, so between an idea for a dialogue and a finished file there is not one decision to make.
- Two people in one file
- A question-and-answer exchange is recorded as one job, with no stitching two files in an editor. In English the only other engine here that does it is VibeVoice 7B.
- A laugh in brackets is performed
- In our recording "(laughs)" became a laugh rather than a word. For a short advert scene that is the detail that separates a dialogue from two sentences read out.
When to pick something else
This is the least predictable engine of the ten. There is no voice parameter, so the timbre differs on every run and two adverts in a series cannot be recorded by the same pair. Asked for one long turn it drifts, and given a single sentence it can truncate it — which is why the sample beside this is a short exchange rather than a monologue. If the dialogue has to be Polish, or longer than a few lines, VibeVoice 7B has named voices, up to four speakers and a 30,000-character limit. If one English voice will do, Kokoro is quicker and repeatable.
See every modelHow to brief this model
The things this engine handles differently from the rest. The habits that hold for every model are collected separately, in the help centre.
- Label the turns with speaker tags and alternate them. Text with no roles marked comes back in one voice, because the tags are the only way to say two people are speaking.
- Stick to a few short turns. On one long utterance the engine drifts, and on a single sentence it can cut it off halfway.
- Do not plan a series on one pair of voices. With no voice field the timbre is drawn afresh each time — for a repeatable read use VibeVoice 7B.
- Write a laugh or a sigh in round brackets inside a turn. Ours performed "(laughs)" rather than reading it.
- Do not type Polish here. The engine is English, and a Polish sentence comes out of it as a run of syllables.
Questions about this model
Other models
- MiniMax Speech 2.8 HDA Polish voice with an emotion you pick and pauses you place.
- MiniMax Speech 2.8 TurboThe same Polish voice as HD, back sooner. For auditioning takes.
- ElevenLabs Multilingual v2It expands "24 h" into "24 godziny" by itself. For copy with numbers in it.
- ElevenLabs Turbo v2.5The shortest wait of the ten. For short adverts and quick takes.
- ElevenLabs v3You write the mood into the middle of the sentence, in square brackets.
- VibeVoice 7BTwo people talking in one file. In Polish.
- KokoroEnglish only, and nothing to configure. For adverts abroad.
- Chatterbox HDThe only engine here that will speak in your own voice.
- Orpheus TTSA laugh or a sigh mid-sentence. English, and slowly.
Try this model on your own product
Sign up without a card and get 40 credits. Switching model is one click, so you can send the same brief through several of them and keep the best frame.