Explore specialized speech-to-text, text-to-speech, stem separation, and audio models. Click any model to launch in Studio or integrate via API.
Multilingual is $0.0052 / min. Diarization is included.
Multilingual is $0.0058 / min. Diarization adds $0.002 / min.
Current English voices. Name the one you want.
Diarization adds $0.02 / hr.
Diarization adds $0.03 / hr.
Diarization adds $0.0025 / min.
Same per-minute price as the file model.
Same character price live or from a file.
Same character price as the file model.
One reference clip, or a saved voice id.
Billed again for every target language.
Turn detection is built in. No diarization.
Send a language hint. No diarization.
Earlier English voices. Name the one you want.
Newer voices. Name the one you want.
Same price as the file model.
Word-level timestamps.
Can translate the audio into English.
25 European languages. Word timestamps.
English only. Word timestamps, plus a blank-penalty slider.
English. Word timestamps from token times.
30 languages. A prompt can pin names and terms.
14 languages. A language code is required.
Partials arrive while you are still talking.
Audio starts as soon as the first words are ready.
Audio keeps arriving while it is still speaking.
Lightweight. A good fit for steady background noise.
Attenuation limit, post-filter, and noise-floor controls.
Same controls as DeepFilterNet 2.
Stereo is billed per channel.
Output format is the only extra choice.
Separates a mix into vocals, drums, bass, and other.
Fine-tuned. Same four stems as HTDemucs.
Adds guitar and piano to the four-stem set.
Vocals against the rest of the mix.
Separates vocals and instrumental, plus drums, bass, and other when you ask.
Six-stem separation: vocals, drums, bass, other, guitar, piano.
Restores muffled or bandwidth-limited speech.
Voices across nine languages, plus speed.
Tags such as [laugh] go in the text.
Exaggeration and CFG weight.
Set the language, plus exaggeration and CFG.
Language, steps, guidance, and speed.
Pick a language and a preset voice.
Native audio is 48 kHz.
Language, preset voice, and an instruct line.
Zero-shot from a short clip.
New text, in the voice of the clip.
Tags such as [laugh] still work.
Same exaggeration and CFG as the speech model.
The clip needs a transcript. Instruct is optional.
Transcript required, unless you use the speaker vector only.
Upload a clip, type what it says, then type new words.
The reference transcript is ignored.
Describe the voice. No reference clip.
The description is tags, not a free-form sentence.
A free-form description, then the words to speak.
Keeps the words from one clip. Uses the voice from another. No transcript.
Continues speech from a prompt clip and more text.
A small managed model is $0.075 per minute. A larger one, such as Claude Sonnet, is $0.163. Your own LLM is $0.065. Your own LLM and voice is $0.050.