About AI Voice Studio

AI Voice Studio is a website with three voice tools on one page: Text to Speech, Voice Clone and Speech to Text. It works in your browser, so there is nothing to install and you don't need a graphics card. All generation happens on servers, not on your computer. You sign in with a Google account, and one credit balance covers all three tools.

What you can do here

• Text to Speech: type up to 3,000 characters with a Standard voice, or up to 5,000 with an HD voice, and get an MP3 you can play in the page and download.
• Voice Clone: upload or record a short voice sample, type up to 1,000 characters, and hear that text in the cloned voice in any of 21 languages.
• Speech to Text: upload an audio or video file of up to 30 MB and 15 minutes and get the full transcript, which you can copy or download as TXT or as SRT subtitles with timestamps.

You can try every tool from the studio on the home page. The voice cloning guide explains how to record a sample that works well.

How it works

Standard voices run on our own servers on Pocket TTS, an open-source speech model. There are 23 preset voices, and each one speaks English, French, German, Spanish, Italian, Portuguese and Dutch. On our server, a 14-second line took about 6.5 seconds to generate when nothing else was queued.

Voice Clone runs entirely on our own servers on two open-source models: Pocket TTS for English, French, German, Spanish, Italian, Portuguese and Dutch, and MOSS-TTS-Nano, a model with 100 million parameters, for the other 14 languages, including Chinese, Japanese and Korean. A new clone is made from your sample for each request; we don't train a custom model of your voice. Your voice sample is not sent to any third-party AI company. Because the models are small, they copy the tone and character of a voice from a short clip, but the result is not identical to the original speaker.

Speech to Text runs on our own servers with the open-source Whisper "small" model. It detects the language automatically, or you can choose it, and it recognizes about 99 languages.

HD voices are the one exception. They come from a commercial cloud voice model with 16 premium voices and 29 languages, so the text you type for HD speech is sent to a third-party voice provider to generate the audio.

We don't add a watermark to any audio, and we don't use your uploads or text to train models.

Open-source credits

These open-source projects make the site possible:

• Pocket TTS by Kyutai, released under CC-BY-4.0, for Standard voices and for Voice Clone in English, French, German, Spanish, Italian, Portuguese and Dutch.
• MOSS-TTS-Nano by the OpenMOSS team, released under Apache-2.0, for Voice Clone in the other 14 languages.
• Whisper by OpenAI, run through faster-whisper, released under MIT, for Speech to Text.

Where the Standard voices come from

The Standard voices are based on recordings released under CC0 or CC-BY-4.0:

• VCTK corpus, University of Edinburgh (CC-BY-4.0)
• Voice actor Alba MacKenna (CC-BY-4.0)
• CML-TTS (CC-BY-4.0)
• Kyutai voice donations and LibriVox readers, via Voice-Zero (CC0)
• Mozilla Common Voice (CC0)

Contact

Questions, feedback or a problem with your account: email [email protected]. You can also see the contact page for what to include in your message.