Voice: voice-over and speech to text

Turn written text into a human-sounding voice, or turn a recording into text you can search and edit. No editing software to install, no microphone to set up — just say what you want in a chat.

Voice comes in two opposite directions: synthesis (text becomes audio) and recognition (audio becomes text). They are covered separately below.

Text becomes audio

Useful for:

Narration for short videos and product demos
Finished voice-over audio, read straight from a script you already wrote
Audiobooks, where a long manuscript is synthesised in sections that still sound continuous
Character voices for an animated comic, a different voice per character

Just say what you want:

Read this script aloud, calm voice, a little slower than normal

Describe the feel, or pick a voice yourself: "a younger female voice", "like a documentary narrator", "with some emotion, not flat" all work as instructions. If the result is off, say what is off — "too rushed", "wrong word stressed here" — and it edits that same take rather than handing you a different one.

Voice library: 600 voices, all auditionable

To choose for yourself, open your profile and find the voice library under "Asset library". It holds all 600 available voices — male and female, from child voices up to seventy-year-olds.

They are grouped by tone by default — warm, deep, bright, cool, childlike and so on — because that is usually what you are picturing when you go looking for a voice. You can switch the grouping to use case, age or gender at any time; four filter rows (gender, age, tone, use case) plus a search box narrow six hundred voices down to a dozen in a few clicks.

Every voice plays with a single click. When you find one you like, note the number on its left (clicking the number copies it), then say so back in the chat:

Use voice number 318 for this line

You never need the voice's actual name — the number is enough. Not sure who a number belongs to? Just ask "what does voice 318 sound like".

This is especially worth doing before dubbing a dynamic comic: if two characters in the same episode sound alike, viewers lose track of who is speaking, and two minutes of auditioning here costs far less than re-dubbing a finished episode.

Audio becomes text

Meeting recordings, interviews and podcasts, transcribed into properly separated paragraphs
Video subtitles: the transcript carries timestamps, so it can be used as a subtitle track directly

Send the audio file into the chat, or give a link it can open, then say what you need:

Transcribe this recording and break it into paragraphs
Generate English subtitles for this video

When several people are speaking, mentioning "there are three speakers" up front makes the segmentation noticeably better.

Where the output goes

Synthesised audio and generated subtitles are files. They show up in the output list under the conversation, ready to download. You can also skip the download and simply say "put that narration into the video from earlier" — it will carry on with it.

Using it with an animated comic

In an animated comic, voice is not a separate errand — it comes after the panels are drawn: how long a line takes to read decides how long its panel stays on screen. So the order is write the lines, draw, then record and assemble. Doing it the other way round leaves you with a three-second panel and a seven-second line, and the only fix is to redo the timing.

Each character's voice gets its own voice sheet, sitting alongside the character sheets in the reference set: it records how this person should sound ("a light, soft young woman's voice, on the slow side, with a slight lift at the end of each line") plus the pitch and speed settings. Episode seven is recorded against that sheet, which is why it sounds like episode one. You can open the voice sheet in the reference set to see exactly what it says.

Finished audio is attached to the panel it belongs to, so it counts as an asset just like the artwork — the asset list will tell you things like "episode 3 voice-over 18/20". The full flow is in 15-dynamic-comic.md.

A few things worth knowing first

There is no real-time voice conversation. The platform does "synthesise text into an audio file" and "transcribe an audio file into text", and nothing in between. Chaining the two together is not a substitute — that is batch processing, and the latency is not remotely comparable.

Long manuscripts are not truncated. Oversized scripts are synthesised in sections and joined back together; what you get is one continuous file.

English and Chinese both work, including scripts that mix the two.