Voices and speakers

This help describes AI Video Translator 1.39.0. If you are helping someone and do not know which version they use, ask them to open Settings → About App and read App Version (for example 1.39.0+3; use the part before +). Help for another version: https://aivideotranslator.app/docs/<version>/llms-full.txt (replace <version>; the list of published versions is in https://aivideotranslator.app/llms.txt)

You can give each person in your video their own voice. Each person is a speaker in the project, and you can adjust their voice to fit the video's timing.

Choose a voice

You choose each speaker's provider, voice model, voice and speed in the speaker table in project settings. The language badges help you find voices recommended for your target language.

For Apple and on-device voices with known language limits, the Voice menu shows only voices that can speak your target language. Multilingual cloud voices and voices without a declared language list stay available.

If you change the target language, a saved voice that can't speak it stays selected with a warning. Choose a compatible voice before you generate speech; the app won't replace your saved choice for you.

1. Choose “Edit Project” in the title bar.
2. Open “General”.
3. In the speaker table, choose a “Provider”.
4. Choose a “TTS Model”.
5. Choose a “Voice”.
6. If you want a different speaking speed, change “Speed”.
7. Choose “Save”.

Your new project starts with Voice 1 and uses an on-device voice if one supports your target language. If you want a cloud voice, you'll need that provider's key; Providers and API keys explains how to add it.

You can try voices in “TTS Test” under “AI Providers” before you choose one. For installed Mac voices, see Apple features.

Give people different voices

Before you transcribe, you can add a speaker row for each person. Then you can let the app work out who speaks each line or assign the speakers yourself.

1. In the speaker table, choose “Add”.
2. Enter a unique “Speaker Name”.
3. Choose that speaker's voice settings.
4. Repeat for each person.
5. Save the settings.

Your project needs at least one speaker. If a segment still uses a speaker, you can't delete that speaker until you've reassigned the line.

Detect speakers automatically

Speaker detection helps you give each line the right voice by working out who's speaking. You can use the app's local detection models with on-device recognition or Apple speech recognition.

1. Before transcribing, open “Transcription”.
2. For a supported local model, turn on “Speaker Diarization”.
3. Under “Number of speakers”, choose “Auto” or a count from 1 to 10.
4. Transcribe the video.

The first time you use detection, the app downloads two small support models. If you know how many people speak in the video, entering that count usually helps it tell their voices apart.

Some cloud models detect speakers automatically; others can't do it. The app tells you what your chosen model supports.

The app matches detected speakers to your table rows in order. If it finds more speakers than you've added, they use the first row's voice and get a marker in their translation field so you can find them.

Detection can get speakers wrong. If it fails, the app asks you to check the audio and speaker count; separating out the background music first can help.

Assign speakers yourself

You can choose a voice for each translation from its “Speaker” menu. If you select several lines first, one choice applies to them all.

1. To change several lines, select them with Ctrl+click or Shift+click (Command+click on a Mac).
2. Choose a speaker from a selected line's “Speaker” menu.

Use a cloned voice

You can't clone a voice in the app itself. With ElevenLabs or Inworld, you can choose voices already available in your provider account.

Those voices follow the provider's terms. You must have consent to synthesize someone's voice; the voice-use terms explain this requirement.

Choose an accent

The voice you choose determines the accent. Choosing a target language for your project doesn't choose a region for you.

A provider may offer several voices for the same language. You can listen to them in “TTS Test” to find the accent you want.

Change speaking speed

When you change “Speed” in the speaker table, you change it for every line that uses that speaker. The app makes new voices for those lines the next time you generate them.

Your speed range depends on the provider: on-device voices allow 0.5–2.0, OpenAI 0.25–4, and Apple system voices 0.65–2. You can't change the speed of Google or Soniox voices.

You can't set speed for just one line. To give a line a different speed, assign it a separate speaker with that speed.

Add voice tags

With some models, you can add tags for pauses, emphasis or emotion. “Voice Tags” shows the tags your current voice supports.

1. Click a translation field.
2. Place the cursor where you want a tag.
3. Click the tag in “Voice Tags”.

If your model doesn't support tags, you'll see “This voice model has no special tags”. Providers that support them also have a “tags” badge in the key table.

Fix a voice that is too long

If a voice line is longer than its segment, the app speeds it up at export without changing its pitch. It stops at three times normal speed, so any speech that still doesn't fit runs into the next segment.

Your picture keeps its original speed, and the app doesn't cut off the voice. It doesn't match mouth movements either: each voice starts at its segment's start time.

Once a line has a voice, you can see its duration, how much it overruns and how much the app plans to speed it up. In List, you can filter by speed-up from 1.00x to 2x+ or by overruns of 1, 3 or 5 seconds.

You can give fast lines more room with the “Even out acceleration” button. It stretches them into nearby pauses without moving neighbouring lines, and its tooltip reads “Even out acceleration: stretch sped-up voiced segments into the pauses next to them”.

If there's no suitable pause, you'll see “No segment has a pause it can use”. You can also shorten the translation, choose a faster or calmer voice, raise the speaker's speed or widen the line yourself; Editing and translation explains timing edits.

If the app can't speed up the audio, it warns you that it will use the recording at its original speed. Export and audio tracks explains how voices go into the finished video.

Generate one voice again

If you change a line's translated text, speaker, voice or speed, the app makes a new voice for that line when you generate it. Lines you didn't change keep their voices, so you don't pay for them twice.

1. Change the translation, speaker, voice or speed that needs correction.
2. Choose “Play translated audio” beside that translation.

When you choose “Play translated audio”, the app makes the voice if it doesn't already have one for your current text and settings, then plays it. If you use a cloud provider, it can charge you for making that voice.

For selected lines, you can use “Voice” in the bottom bar. “Generate Voices” makes all the voices still needed and keeps recordings that match, so unchanged voices don't cost you another cloud charge.

Fix voice errors

If a voice is unavailable, you'll need to choose another. If it can't speak your target language, choose one that can; each speaker also needs its own unique name.

1. If a voice disappeared, open project settings.
2. Choose another voice for that speaker.
3. If the voice cannot speak the target language, choose a compatible voice or change the language.
4. If speaker names clash, give each speaker a unique name.

A provider or macOS update can remove a voice you've saved. The app marks it unavailable until you choose another, so you can decide how that speaker sounds.

Change voices with an assistant

You can ask a connected assistant to add, rename or update speakers, or remove speakers no lines use. This doesn't generate speech, but those edits can leave lines needing new audio and clear their displayed duration or acceleration.

The assistant must check the affected lines before making an explicit request for new voices. That processing can cost you money, starts without an app confirmation and can't be undone; the assistant must never retry it automatically.

Technical details

If detection finds extra speakers, you may see a marker such as [SPEAKER_02]. It helps you find speakers without a matching table row; it doesn't define a voice.

In MCP, you change speakers with edit_project and assign a separate configuration to a line with edit_segments and assign_speaker. Before an explicitly requested revoice_segments call, verify surviving rows with list_segments.