Your presenter's voice, on demand.
Six seconds of clean audio is enough to build a voice that will read any script you write, in Arabic, English or French. Useful at two in the morning when the bulletin needs a track and nobody is in the booth.
Six seconds in, any script out, in three languages. Open any stage for what it needs.
A voiceover is thirty seconds of work that requires a person, a room and a schedule. So the overnight update waits until six, the online version of the package goes out mute, and the second-language edition never gets made at all — not because anyone decided against it, but because the booth was locked.
Voice Studio is for the reads that never justified waking somebody up. It is not a replacement for your presenters — a channel's identity lives in the voice, and nobody should want that automated. It is for the timing track on a rough cut, the overnight web version, the fourth language nobody has staff for, and the two in the morning when a story will not wait until the shift changes.
10 things it does. Open any one for how it works.
A short clean sample builds a voice that will then read anything. A minute of good audio makes it noticeably better, but you do not need a recording session to start.
Clean matters more than long. Six seconds from a quiet studio beats two minutes recorded in a newsroom with a television on in the background, because whatever is in the sample becomes part of the voice.
Modern Standard Arabic with the rhythm and stress of broadcast delivery, not the flattened cadence that makes machine reading obvious within three words.
The same voice reads Arabic, English and French. A package can carry a second-language track without finding a second presenter.
Every sentence is generated as its own piece as well as one continuous read, so fixing a single mispronounced name means regenerating that line rather than the whole script.
Blend several speakers into one voice that resembles none of them — useful when you want a consistent sound for routine material without attaching it to a named presenter.
This also solves a practical problem. A synthetic voice built from a named presenter raises a question every time it is used; a house voice that is nobody's raises none, and can carry the overnight web reads without implying that person was in the building.
A read generated for a script arrives in the package with the timings already right, so a rough cut has real durations rather than estimates.
Use the generated track for timing, then drop the presenter's read over it when the booth frees up. The edit does not have to be rebuilt around the new audio.
Past reads stay in a library with the script that produced them, so a corrected line can be regenerated months later in the same voice.
Where you have hours of a presenter's archive, a dedicated voice can be trained on it for a closer match than a short sample allows.
Clean audio still matters more than the amount of it, but with hours of a presenter's archive behind it the match is closer than any short sample can give.
Building a voice, mixing one and using one are separate permissions. A presenter's voice can exist in the system without every producer being able to make it say things.
Upload a short clean sample of whoever is going to read, or blend several into a house voice. This is a five-minute job you do once.
Or send it straight from a package in TV Studio, in which case the read comes back attached to the right scenes.
A name pronounced wrong is one line to regenerate, not a whole take. Adjust the spelling in the script and ask for that sentence again.
The track drops onto the timeline with the music already ducking under it. Replace it with a presenter later if the booth opens up.
A story moves at two in the morning and the web team wants a forty-second video by three. The booth opens at seven. The house voice reads it, the package goes up before the morning commute, and the presenter records the broadcast version when they come in.
A producer building a package needs to know whether the script actually runs to ninety seconds before they cut pictures to it. A generated read answers that in under a minute, and the pictures are cut to real durations from the start rather than being re-trimmed after the presenter records.
The package already exists in Arabic. A translated script and the same voice produce a French track, and a version of the story reaches an audience the channel was not previously serving — without hiring for it.
Nothing is built from a presenter's audio without somebody deciding to build it. The library is what you put in it.
Building, blending and using are separate rights. A named presenter's voice can be restricted to the people entitled to use it.
Fix one sentence and leave the rest of the read exactly as it was, including anything you had already placed on a timeline.
Every place a generated track can go, a recorded track can go instead. Nothing in the platform requires it.
A voice comes from audio somebody supplied deliberately. There is no route from footage in the archive to a usable voice without a person making that choice.
Generated audio is labelled as generated throughout the platform, so nobody downstream mistakes it for a booth recording.
Dubbing a speaker into another language while keeping their own voice is not built. Where it appears in the interface it reports honestly that it is unavailable.
A read is a file on a timeline. What happens to the package after that is somebody's decision.
This is not aimed at replacing the voice your audience knows. The value is in the work that currently does not get done at all — the overnight update that waits until the shift changes, the online cut that goes out mute, the second-language edition nobody has staff for, the timing track a producer needs before they can cut pictures. Those are not jobs anyone is losing. They are jobs nobody was doing, because a thirty-second read required a person, a room and a schedule.
That is an editorial policy question rather than a technical one, and most channels start by using it off-air — timing tracks, online cuts, overnight updates — while keeping broadcast reads with their presenters. The system supports either; the decision is yours to make and to disclose.
You should treat it that way, and the permission model is built to support it — building a voice, and using one, are separate rights that can be held by different people. A blended house voice avoids the question entirely for routine material.
Not in the speaker's own voice — that is not built, and where it appears in the interface it says so rather than producing something misleading. You can generate a translated read in one of your own voices, which is a different thing and honestly labelled.
Quiet matters far more than long. Six seconds recorded properly gives a better result than two minutes with room noise, because everything in the sample — including the air conditioning — becomes part of the voice.
The useful demonstration is the one that runs on your footage and your wire feeds. Tell us how your desk works and we will show you the parts that matter — and say so where we do not help.