A meeting starts in English, someone answers in Spanish, and another speaker casually drops a French product name into the middle of a sentence. That kind of conversation is normal for international teams. For older speech-to-text tools, though, it can turn into a wonderfully creative mess.
The Gemini 3.5 Transcribe supported languages list covers more than 85 language and locale options. The model can automatically detect spoken language, follow some language changes during a conversation, and accept a specific language code when you already know what’s being spoken.
There’s one detail worth understanding early: support inside the transcription model doesn’t always mean the same language is available in every Gemini product. The Gemini API offers the broadest language coverage, while features on Mac, Android, and other consumer products may have smaller language lists.
How Many Languages Does Gemini 3.5 Transcribe Support?
Google says Gemini 3.5 Transcribe can automatically recognize speech across more than 85 language and locale settings.
A locale is slightly more specific than a language. English spoken in the United States, Great Britain, and India appears under separate BCP-47 codes:
- English (United States):
en-US - English (Great Britain):
en-GB - English (India):
en-IN
The same distinction applies to Portuguese, Spanish, Bengali, Punjabi, and a few other languages. This matters because pronunciation, spelling, vocabulary, and number formatting can vary between regions.
Google’s official Gemini audio transcription documentation contains the current technical language list and corresponding codes.
Full Supported Language List by Region
The following list is organized by region to make it easier to scan. Google may change language coverage as the model develops, so check the official documentation before building a production workflow around a particular language.
English and Major Regional Variants
- English (United States)
- English (Great Britain)
- English (India)
- Indian English
For most American podcasts, calls, interviews, and voice notes, en-US is the sensible language hint. British recordings should use en-GB, while en-IN is intended for Indian English.
Western and Northern European Languages
- Catalan
- Danish
- Dutch
- Finnish
- French
- Galician
- German
- Icelandic
- Italian
- Norwegian
- Portuguese (Brazil)
- Portuguese (Portugal)
- Spanish (Latin America)
- Spanish (United States)
- Swedish
Spanish deserves a quick note. Google currently lists es-419 for Latin American Spanish and es-US for Spanish spoken in the United States. If you’re transcribing a recording made for a US Hispanic audience, es-US may be the more suitable starting point.
Central and Eastern European Languages
- Belarusian
- Bosnian
- Bulgarian
- Bulgarian (Aromanian)
- Croatian
- Czech
- Estonian
- Greek
- Hungarian
- Latvian
- Lithuanian
- Macedonian
- Polish
- Romanian
- Russian
- Serbian
- Slovak
- Slovenian
- Ukrainian
Several languages in this group share words or similar sounds, especially in short clips. Giving the API the expected language code can reduce confusion when the model has only a few seconds of audio to examine.
South Asian Languages
- Assamese
- Bengali (Bangladesh)
- Bengali (India)
- Gujarati
- Hindi
- Kannada
- Malayalam
- Marathi
- Nepali
- Oriya
- Punjabi
- Punjabi (Gurmukhi script)
- Sindhi (Arabic script)
- Telugu
Regional variants matter here too. Bengali has separate codes for Bangladesh and India, while Punjabi appears with both a general Indian code and a Gurmukhi-script option.
Urdu isn’t explicitly included in Google’s current published language table. That doesn’t prove the model will fail to recognize every Urdu recording, but it does mean you shouldn’t promise reliable Urdu transcription without testing it carefully first.
East and Southeast Asian Languages
- Burmese
- Cantonese (Traditional)
- Cebuano
- Central Khmer
- Filipino
- Indonesian
- Japanese
- Javanese
- Korean
- Malay
- Mandarin Chinese (Simplified)
- Mongolian
- Thai
- Vietnamese
Mandarin Chinese and Cantonese are treated separately. For Mandarin, Google lists Simplified Chinese with the code cmn-Hans-CN. Traditional Cantonese uses yue-Hant-HK.
Choosing the correct option matters when names, locations, and written characters need to remain accurate.
Middle Eastern, Caucasian, and Central Asian Languages
- Arabic (Egypt)
- Armenian
- Azerbaijani
- Farsi
- Georgian
- Hebrew
- Kazakh
- Kyrgyz
- Tajik
- Turkish
- Uzbek
Google currently names Arabic (Egypt) rather than providing a single universal Arabic option. Arabic varies considerably across countries, so recordings from other regions should be tested before the model is used for important transcripts.
African Languages
- Afrikaans
- Amharic
- Hausa
- Kabuverdianu
- Lingala
- Swahili (Kenya)
Support for these languages is useful, but the quality of the recording still matters. Background noise, overlapping voices, weak microphones, and strong regional variation can affect the result even when the language appears on the supported list.
Common Language Codes You May Need
BCP-47 codes tell the transcription system which language and regional form to expect.
| Language | BCP-47 code |
|---|---|
| Arabic (Egypt) | ar-EG |
| Bengali (Bangladesh) | bn-BD |
| Bengali (India) | bn-IN |
| Cantonese (Traditional) | yue-Hant-HK |
| English (Great Britain) | en-GB |
| English (India) | en-IN |
| English (United States) | en-US |
| French | fr-FR |
| German | de-DE |
| Hindi | hi-IN |
| Indonesian | id-ID |
| Italian | it-IT |
| Japanese | ja-JP |
| Korean | ko-KR |
| Mandarin Chinese (Simplified) | cmn-Hans-CN |
| Portuguese (Brazil) | pt-BR |
| Portuguese (Portugal) | pt-PT |
| Russian | ru-RU |
| Spanish (Latin America) | es-419 |
| Spanish (United States) | es-US |
| Swahili (Kenya) | sw-KE |
| Thai | th-TH |
| Turkish | tr-TR |
| Vietnamese | vi-VN |
You don’t have to memorize these codes. Most people won’t. Save the official language table and copy the required code when setting up a transcription request.
How Automatic Language Detection Works
Gemini 3.5 Transcribe detects the spoken language automatically when you don’t provide a language hint.
For an API request, automatic detection can be used by leaving language_codes out of the configuration or passing an empty list:
{
"transcription_config": {
"language_codes": []
}
}
This is handy when recordings arrive from different countries and you don’t know the language beforehand.
But automatic detection isn’t magic. A three-second clip containing a person’s name and one short sentence gives the model far less context than a clear, two-minute recording. Similar languages, heavy noise, and unusual accents can also make detection harder.
When you know the language, provide the code. It removes one decision from the model and may improve the result.
How to Set a Language Hint
Suppose you’re transcribing a customer interview recorded in American English. The language configuration would look like this:
{
"transcription_config": {
"language_codes": ["en-US"]
}
}
For Latin American Spanish, replace the code with es-419:
{
"transcription_config": {
"language_codes": ["es-419"]
}
}
The language hint doesn’t translate the audio. It simply tells the transcription model which spoken language or locale it should expect.
If you need help setting up the model itself, follow our step-by-step Gemini 3.5 Transcribe guide. It explains the available access methods, recorded-audio model, live transcription option, and basic API workflow.
Can It Handle More Than One Language in a Recording?
Yes. Google says the model can handle code-switching, which happens when someone moves between languages during the same sentence or conversation.
Picture a bilingual sales call:
“The delivery is confirmed for Friday, pero necesitamos la dirección completa.”
A multilingual transcription system should recognize the switch from English to Spanish without treating the Spanish words as badly pronounced English.
This can be genuinely useful for:
- International team meetings
- Bilingual interviews
- Customer-support calls
- Travel videos
- Classroom discussions
- Podcasts with overseas guests
- Voice notes containing foreign names or phrases
Still, review mixed-language transcripts carefully. A model may understand the main language perfectly but miss a short phrase, regional expression, or proper name from the second language.
Language Support and Product Availability Aren’t the Same
This is where people often get confused.
Gemini 3.5 Transcribe may support more than 85 locales at the model level, but the feature you’re using might have a narrower list.
For example:
- The Gemini API provides access to the model’s broader transcription settings.
- Gemini voice features on macOS may initially support fewer languages.
- Gboard features can depend on the phone model, region, app version, and selected keyboard language.
- A feature may appear gradually rather than reaching every account on the same day.
So if a supported language doesn’t appear inside a consumer app, that doesn’t necessarily mean the underlying transcription model lacks it. The product interface may simply not offer it yet.
Students already using Google’s education tools can also read our Gemini Student Hub guide for a clearer look at Gemini’s study-focused features.
Does Accent Affect Transcription Quality?
Language support and accent support aren’t quite the same thing.
Google says the model is designed to handle diverse accents, background noise, and multilingual conversations. Even so, accuracy can change depending on:
- Microphone quality
- Recording distance
- Speaking speed
- Regional pronunciation
- Background music
- Several people talking at once
- Uncommon names
- Technical vocabulary
- A very short audio sample
An American English code doesn’t mean every American accent will produce identical results. A clean recording from a quiet room will usually be easier to process than a phone call made inside a busy train station.
And honestly, names are where I’d be most cautious. A transcript can look polished and still spell one person’s name incorrectly.
Use Custom Vocabulary for Names and Technical Terms
Language selection tells Gemini what language to expect. Custom vocabulary helps it recognize specific words inside that language.
You can provide terms such as:
- Employee names
- Company names
- Product names
- Medical terminology
- Legal phrases
- Industry abbreviations
- Unusual place names
Google allows up to 1,000 custom vocabulary terms, though its documentation says the best results are generally achieved with a focused list of up to 100 terms.
Don’t fill the list with ordinary words. Add the terms the model is genuinely likely to misunderstand.
There’s also a technical limitation: custom vocabulary can’t currently be combined with speaker diarization or word-level timestamps in the same request. Decide which feature matters most before configuring the transcription.
Practical Ways to Improve Multilingual Transcripts
Record Clean Audio
Keep the microphone close enough to capture speech clearly. Avoid loud music, fans, traffic, and room echo where possible.
A supported language can still produce a poor transcript when the original recording is difficult to hear.
Provide the Language Code When You Know It
Automatic detection is convenient, but a correct language hint gives the model useful context from the beginning.
This is especially helpful for short recordings and languages with similar vocabulary or pronunciation.
Keep Regional Variants in Mind
Use en-US for American English rather than choosing English at random. The same logic applies to Brazilian and European Portuguese or Latin American and US Spanish.
Add Important Names to Custom Vocabulary
If a podcast guest has an unusual surname, add it before processing the recording. Do the same for brand names, medical terms, software products, and abbreviations.
Review Language Changes Manually
Code-switching support can save time, but don’t publish a bilingual transcript without listening to the sections where the language changes.
Those short transitions are often where errors hide.
Test Before Processing a Large Archive
Take two or three representative recordings and compare the transcripts first. Use clips with different speakers, accents, and noise levels.
A quick test is much cheaper than discovering a recurring mistake after processing hundreds of files.
Supported Audio Formats
Language support won’t help if the uploaded audio format isn’t accepted.
Gemini 3.5 Transcribe currently supports common formats including:
- WAV
- MP3
- AIFF
- AAC
- OGG
- FLAC
- MPEG
- M4A
- L16
- Opus
- ALAW
- MULAW
- WebM
For longer recordings, Google recommends uploading the file through its Files API and passing the returned file location to the transcription model.
Frequently Asked Questions
Does Gemini 3.5 Transcribe detect languages automatically?
Yes. If no language code is supplied, the model can automatically identify the spoken language. You can also pass an empty language_codes list to use automatic detection.
Can it transcribe English and Spanish in the same recording?
Yes, the model is designed to handle code-switching between languages. Results should still be checked around the points where the speaker changes language.
Does Gemini 3.5 Transcribe support Urdu?
Urdu isn’t explicitly listed in Google’s current official supported-language table. Test Urdu recordings before relying on the model, and don’t assume production-level support until Google lists it clearly.
Which English variants are supported?
Google lists English for the United States, Great Britain, and India. Their primary codes are en-US, en-GB, and en-IN.
What code should I use for Spanish?
Use es-419 for Latin American Spanish or es-US for Spanish spoken in the United States.
Can the model translate speech into another language?
Transcription and translation are different jobs. Gemini 3.5 Transcribe converts speech into written text. A separate translation step may be required when you want the final text in another language.
Do I have to enter a language code?
No. The model can detect the language automatically. Supplying the correct code is useful when you already know the language and want to give the model clearer guidance.
Will every supported language work in Gboard or the Gemini app?
Not necessarily. Consumer features can have different device, region, account, and language restrictions. The model’s API language list is broader than what may appear in a particular app.
Choose the Language Setting Before You Press Upload
The Gemini 3.5 Transcribe supported languages list is broad enough for international meetings, multilingual interviews, podcasts, lectures, and customer calls. Automatic detection makes the first test easy, while BCP-47 hints give you more control when the language is already known.
But don’t judge a transcript only by how polished it looks. Listen to names, numbers, regional expressions, and moments where the speaker changes language. A clean paragraph can still contain one confidently written mistake.
Start with a short sample, use the correct locale code, and add only the specialist terms that matter. That small amount of preparation can save a surprising amount of correction later.
