Multilingual Guide

Multilingual Guide

Vayu supports 99 languages via the Whisper multilingual models.

Language Detection

By default, Vayu auto-detects the language from the first 30 seconds of audio:

result = whisper.transcribe("audio.mp3")
print(result["language"])  # e.g., "en"

For better accuracy, specify the language explicitly:

result = whisper.transcribe("audio.mp3", language="fa")

Supported Languages

Code	Language	Code	Language	Code	Language
en	English	ja	Japanese	uk	Ukrainian
zh	Chinese	pt	Portuguese	el	Greek
de	German	tr	Turkish	ms	Malay
es	Spanish	pl	Polish	cs	Czech
ru	Russian	ca	Catalan	ro	Romanian
ko	Korean	nl	Dutch	da	Danish
fr	French	ar	Arabic	hu	Hungarian
it	Italian	sv	Swedish	ta	Tamil
hi	Hindi	id	Indonesian	no	Norwegian
fa	Persian	fi	Finnish	th	Thai
vi	Vietnamese	he	Hebrew	ur	Urdu

Full list of all 99 languages

Code	Language	Code	Language
hr	Croatian	bg	Bulgarian
lt	Lithuanian	la	Latin
mi	Maori	ml	Malayalam
cy	Welsh	sk	Slovak
te	Telugu	lv	Latvian
bn	Bengali	sr	Serbian
az	Azerbaijani	sl	Slovenian
kn	Kannada	et	Estonian
mk	Macedonian	br	Breton
eu	Basque	is	Icelandic
hy	Armenian	ne	Nepali
mn	Mongolian	bs	Bosnian
kk	Kazakh	sq	Albanian
sw	Swahili	gl	Galician
mr	Marathi	pa	Punjabi
si	Sinhala	km	Khmer
sn	Shona	yo	Yoruba
so	Somali	af	Afrikaans
oc	Occitan	ka	Georgian
be	Belarusian	tg	Tajik
sd	Sindhi	gu	Gujarati
am	Amharic	yi	Yiddish
lo	Lao	uz	Uzbek
fo	Faroese	ht	Haitian Creole
ps	Pashto	tk	Turkmen
nn	Nynorsk	mt	Maltese
sa	Sanskrit	lb	Luxembourgish
my	Myanmar	bo	Tibetan
tl	Tagalog	mg	Malagasy
as	Assamese	tt	Tatar
haw	Hawaiian	ln	Lingala
ha	Hausa	ba	Bashkir
jw	Javanese	su	Sundanese
yue	Cantonese

Language Name Aliases

Vayu accepts both language codes and common name variations:

# All equivalent
whisper.transcribe("audio.mp3", language="zh")
whisper.transcribe("audio.mp3", language="chinese")
whisper.transcribe("audio.mp3", language="mandarin")

# More aliases
# "burmese" → my, "valencian" → ca, "flemish" → nl
# "haitian" → ht, "castilian" → es, "panjabi" → pa
# "moldavian" → ro, "sinhalese" → si, "pushto" → ps

RTL Languages (Persian, Arabic, Hebrew, Urdu)

Vayu transcribes RTL languages correctly. The text output is in logical order (not visual order), which is how most text systems expect it:

whisper = LightningWhisperMLX(model="large-v3", batch_size=6)

# Persian
result = whisper.transcribe("persian_audio.mp3", language="fa")
print(result["text"])  # سلام، حال شما چطور است؟

# Arabic
result = whisper.transcribe("arabic_audio.mp3", language="ar")

Tips for RTL:

• Use large-v3 for best accuracy on RTL languages
• Specify the language explicitly — auto-detection sometimes confuses Arabic/Persian/Urdu
• SRT/VTT output may need RTL markers depending on your video player

CJK Languages (Chinese, Japanese, Korean, Cantonese)

CJK languages have special word splitting behavior for word timestamps:

result = whisper.transcribe("chinese_audio.mp3", language="zh", word_timestamps=True)

# Word timestamps use character-level splitting for CJK
for word in result["segments"][0]["words"]:
    print(f"[{word['start']:.2f}] {word['word']}")

CJK, Thai, Lao, and Myanmar use unicode-boundary splitting instead of whitespace splitting, since these languages don't use spaces between words.

Translation to English

Translate any of the 99 languages to English:

# French → English
result = whisper.transcribe("french.mp3", language="fr", task="translate")

# Japanese → English
result = whisper.transcribe("japanese.mp3", language="ja", task="translate")

vayu french.mp3 --language fr --task translate

Translation uses the same model — no separate translation model is needed.

Model Selection for Multilingual

Scenario	Recommended Model
English only	distil-large-v3 or .en variants
Common languages (top 20)	distil-large-v3
Low-resource languages	large-v3 (best multilingual accuracy)
Translation to English	large-v3
Language detection	large-v3 (most reliable detection)

The .en model variants (e.g., tiny.en, base.en) are English-only and cannot transcribe other languages.

Multi-Language Audio

If your audio contains multiple languages (e.g., code-switching), Vayu processes it in 30-second chunks. Each chunk detects language independently when language=None:

# Auto-detect language per chunk
result = whisper.transcribe("multilingual_audio.mp3")

For better results with mixed-language audio, use large-v3 and let auto-detection handle each segment.