AI voiceover voices by language

Marketing pages for AI voiceover tools all say roughly the same thing: a hundred languages, two hundred voices, sometimes more. None of them answers the question a person actually has, which is what a Marathi video is going to sound like, and whether the Malay one will sound the same as the German one. Those two answers live in a lookup table in the source, so here it is.

How a voice is chosen from the language

You give the generator a language. The picker offers 32 common ones, most with their native name beside them, the field is free text so anything you type is passed through, and leaving it empty lets the model match your brief.

Three things then happen. The model writes the script in that language. It picks a voice style and a voice gender from the tone of the brief. A lookup maps your language name to a single Google locale, and style, gender and locale are combined into one voice id.

That last step is the one that matters, because Chirp 3 HD voice names are locale locked. A voice id looks like es-ES-Chirp3-HD-Puck. The character name is a style, not a person: Puck in Spanish and Puck in Japanese are different recordings that share a temperament. You cannot carry one narrator across languages, and no tool built on this catalogue can.

The mapping is one language name to exactly one locale, which means somebody chose an accent on your behalf. Portuguese resolves to pt-BR, not European Portuguese. Spanish resolves to es-ES, not Latin American Spanish. Urdu goes to ur-IN rather than ur-PK, Norwegian to nb-NO, and English to en-US, even though Google publishes en-GB, en-AU and en-IN as separate Chirp 3 HD locales. If your audience is Mexican, Portuguese, Pakistani or British, the script will be right and the accent will not, and the fix is to override the voice on each scene rather than to change the language.

The 56 locales

The table at the foot of this post is the whole mapping, read from the source on 1 September 2026. There are 56 language names and 56 distinct locales, so no two languages share a voice pool.

Two honest notes. First, 56 is smaller than the “100+” you see elsewhere, which usually counts locale variants, or counts every language the text model can write rather than every language that has a voice. Second, the list is moving: an internal check in August recorded 31 mapped languages and the current source has 56, with the Nordic, Baltic, Balkan and South Indian entries all recent.

Anything not in the table gets no narration. There is no silent substitution of an English narrator reading foreign words, which is a deliberate decision in the code and the right one, because an English voice reading Vietnamese is worse than no voice.

The eight voices by style

The automatic pick is four styles crossed with two genders.

Style	Female	Male	Used by default for
Energetic	Laomedeia	Puck	Promotional and UGC videos
Instructional	Erinome	Charon	Tutorial videos
Warm	Sulafat	Umbriel	Chosen when the brief is calm, premium or sensitive
Neutral	Aoede	Iapetus	The fallback when no style is resolved

Those eight are a selection, not the catalogue. Google’s own Chirp 3 HD page lists around thirty voice names, all of them moons, and every one of them is reachable from the voice dialog once a scene exists. The eight are simply the ones the director is allowed to choose without being asked.

One rule overrides everything: a tutorial can never come back energetic. If the model asks for the advertisement voice on a walkthrough, the request is replaced with instructional before the voice id is assembled. That is worth knowing if you have ever wondered why a how-to video refuses to sound upbeat. If you are writing one, the tutorial script guide covers the pacing that voice is built for.

Where quality drops: the four gaps

Three locales have no Chirp 3 HD at all. Malay (ms-MY), Filipino (fil-PH) and Chinese Traditional (cmn-TW) fall back to WaveNet, Google’s older neural tier. WaveNet is perfectly intelligible and noticeably flatter, with less breath and less variation across a sentence. There are two voices in each, not eight, so the style choice does nothing in those three languages.

Russian has three of the eight. Laomedeia, Erinome, Umbriel, Sulafat and Iapetus do not exist in ru-RU, so any request for them is swapped for Charon or Aoede. In practice a Russian promotional video with a female narrator comes back neutral rather than energetic, because Aoede is the substitute. Male Russian promos still get Puck.

Persian is offered and not mapped. It sits in the language picker, the script is written in Persian, and no narration is produced, because there is no locale for it. The captions still carry the copy.

Everything outside the 56 behaves like Persian. The field accepts free text, so you can type Welsh or Zulu and get a Welsh or Zulu script with no voice on it.

One more thing, flagged honestly. Six of the mapped languages, Afrikaans, Amharic, Basque, Catalan, Galician and Icelandic, do not appear on Google’s published Chirp 3 HD locale list as it stood on 1 September 2026. The mapping expects them to be there. Generate a single clip in one of those six before you build a whole video on it.

Captions by script

Captions are burned into the exported file: a single dark pill centred at the bottom, each word lighting up as it is spoken, sitting 120 pixels above the bottom edge on a vertical video, 56 on a square one and 44 on a landscape one, capped at 72% of the frame width.

Two behaviours are worth knowing before you commit to a language.

Direction is detected from the characters, not the language setting. If the copy contains Hebrew or Arabic script, the whole composition switches to right to left and text authored left aligned mirrors to the right edge. That is why Persian copy still lays out correctly with no voice attached: the script triggers it, not the setting. The W3C makes the same point about the web, that base direction has to be declared explicitly and cannot be inferred from the language.

Script fonts load themselves. The design fonts cover Latin plus some Cyrillic and Greek. When the copy contains anything else, the matching Noto Sans face is pulled in behind them and the render waits for it, which is what stops a Tamil or Amharic video exporting as rows of empty boxes. Twenty two faces are wired up this way, covering Japanese, Korean, Simplified Chinese, Arabic, Hebrew, Thai, Devanagari, Bengali, Gurmukhi, Gujarati, Tamil, Telugu, Kannada, Malayalam, Sinhala, Ethiopic, Khmer, Lao, Myanmar, Georgian, Armenian, and a general face for Greek and Cyrillic.

In practice, Chinese, Japanese and Thai captions have coarser word timing than English ones, because those scripts do not put spaces between words and the highlight has less to work with.

Speed, delay and per-scene overrides

Every narration clip is edited on its own scene, not on the video. Each one carries its own text, language, voice, a delay in seconds before the line starts, a speed of 0.85x, 0.9x, 1x, 1.1x or 1.25x, and a caption toggle that is on by default. A scene built from several screenshots can carry a separate line and a separate voice per image.

Slowing a line to 0.85x is the cheapest fix for a language that reads faster than English. The speaking time calculator will tell you roughly how long a script runs before you generate anything.

Trying a language before you commit

Generate one scene, not the video. Open its voice dialog, set the language, press generate, and listen to the clip. Ten seconds of audio tells you more than any language list, especially in the three WaveNet locales where the difference is audible rather than theoretical.

Do it before re-narrating a finished video, because a language change means every clip is synthesised again, and a video that needs new narration leaves the free export path that a 30 second cut would otherwise sit on.