Text to Speech and Speech to Text
Turn a script into narration with AI Text to Speech, or a recording into an editable, timed transcript with Speech to Text: limits, voices, costs and export formats for both.
Two tools convert between words and spoken audio. AI Text to Speech reads your script aloud in a preset voice. Speech to Text writes down what a recording says, with timings you can export as subtitles.
Text to speech
- Type or paste up to 5,000 characters into Text.
- Pick a Voice. The picker holds 67 preset voices, each with a recorded preview, so you can compare them before you spend credits.
- Leave Language on Auto-detect, or choose a language when a short passage or a name could be read in the wrong one.
- Optionally open Advanced to set the delivery:
| Setting | What it changes | Range |
|---|---|---|
| Speed | How fast the voice speaks | 0.7–1.2, 1.0 default |
| Stability | How steady the delivery stays from line to line | 0–1 |
| Similarity | How closely the result follows the preset voice | 0–1 |
| Style | How much expression the voice adds | 0–1 |
The voice model is Turbo v2.5. Speech costs 10 credits per 1,000 characters, rounded up to whole credits: an 89-character line costs 1 credit, 1,001 characters cost 11, and a full 5,000 cost 50. The button shows the cost for your text before you start. The narration plays and downloads as an MP3 from Results.
For a script longer than 5,000 characters, split it at paragraph breaks and generate each part with the same voice and settings. Listen to the joins before you use them.
Text to speech speaks with preset voices only. To sing, use the AI Music Generator; to make a voice of your own, see Voice Models and Covers.
Speech to text
- Choose the source: upload an audio file, record with your microphone, pick a song from My Music, or start from one of the example recordings on the page. Recordings can be up to 50 MB and 15 minutes. One example also shows the transcript this tool returned for it, so you can see the output before you spend credits.
- Leave Language on Detect automatically, or pick the language spoken from the list of 100 languages the model transcribes.
- Start the transcription. One run costs 4 credits.
The transcript arrives as timed segments, with word timings where they were available. It does not label different speakers, so check recordings with several voices carefully. Clear speech with little background sound gives the most accurate result; proofread names, numbers and technical terms.
Correct and export
Edit any line in the transcript editor, then copy the text or download it:
| Format | Use it for |
|---|---|
| TXT | Notes, scripts and documents |
| SRT | Subtitle editors and most video platforms |
| VTT | Captions on the web |
| JSON | Software that reads the language, text and timings as structured data |
Your corrections stay in the editor and go into what you copy or download; they do not change the original transcription in your history. Download the corrected copy before you leave the page. Correcting a segment's words removes that segment's word timings and keeps its start and end.
The music video editor uses the same transcription and review controls to put timed lyrics on screen; see Music Video Types and Export.
Related: Credits and Plans · Errors and What to Do