Indian languages often arrive typed in Latin letters, in chat messages and in
LLM replies written to sound like chat: aapka EMI kal due hai, please pay karo.
By default the engine treats any word in Latin letters as English, so aapka
and karo get English pronunciations.
Send "romanized": true and the engine reads those sentences as the language
in lang instead. The field exists on both endpoints, and it is off unless you
ask for it.
Languages
Romanized input works for the nine Indian languages: bn, gu, hi, kn,
ml, mr, or, ta and te. That is every language in
Voices and languages except en.
Set lang to the language the text is written in. Latin letters do not say
which language they spell, so the engine goes by lang. Send romanized Tamil
with "lang": "hi" and the Hindi model reads it, which comes out wrong.
What changes
The engine looks at the text one sentence at a time.
| Sentence | With "romanized": true |
|---|---|
Aapka EMI kal due hai. |
The engine reads it as Hindi. EMI and due stay English. |
आपका बैलेंस 500 रुपये है। |
Unchanged. A sentence with any native-script letters reads the same as without the flag. |
Your payment is due tomorrow. |
Unchanged. The engine recognises an English sentence and leaves it as English. |
Inside a romanized sentence the engine reads the language’s own words in the
language. English words that belong in the sentence, like EMI, payment or
loan, keep their English pronunciation.
Pronunciation overrides mostly still apply. A single word you have given a
pronunciation in g2p_overrides, or g2p_overrides_json on the WebSocket,
keeps that pronunciation when you type it in capitals, like KOTAK. Typed in
lowercase, a key that is also a common word of the language goes through the
romanizer instead. So do the most common words, such as ka, se and hai in
Hindi, however you type them.
If the romanizer fails on some text, the engine reads that text the same as without the flag. The request does not fail because of it.
Turn it on when the text may contain romanized sentences, such as chat messages, LLM replies or message templates written in Latin letters. Text that is all native script or all English sounds the same either way.
Billing does not change. You pay for the characters you send, counted by Unicode code point.
Over HTTP
Add romanized to the request body of
POST /tts/bytes.
curl -X POST https://tts.navana.ai/tts/bytes \
-H "X-API-Key: $BODHI_API_KEY" \
-H "Content-Type: application/json" \
-d '{"text":"Aapka EMI kal due hai, please 3 baje tak payment kar dijiye.","lang":"hi","output_format":"8000:pcm16","romanized":true}' \
--output speech.pcm --dump-header -import os, requests
response = requests.post(
"https://tts.navana.ai/tts/bytes",
headers={
"X-API-Key": os.environ["BODHI_API_KEY"],
"Content-Type": "application/json",
},
json={
"text": "Aapka EMI kal due hai, please 3 baje tak payment kar dijiye.",
"lang": "hi",
"output_format": "8000:pcm16",
"romanized": True,
},
timeout=120,
)
response.raise_for_status()
print(response.headers["X-Romanized-Applied"]) # "true"The response body is raw audio, the same as any other /tts/bytes call. See
Non-streaming synthesis to turn it into a WAV.
The X-Romanized-Applied header
The response carries this header when the request included romanized, as
true or false. A request without the field gets no header, so existing
clients see no change.
| Value | Meaning |
|---|---|
true |
You sent true and the server supports romanized input for lang, so the text went through the romanizer. Native-script and English sentences in it still come out unchanged. |
false |
You sent false, or the server has no romanized support for lang. The engine read the text the same as without the flag. |
Over the WebSocket
Add romanized to the hello frame of
wss://tts.navana.ai/v1.
{
"type": "hello",
"lang": "hi",
"output_format": "24000:pcm16",
"romanized": true,
"auth_token": "bd_xxxxxxxxxxxx.xxxxxxxx…"
}Like every hello field, it is fixed for the session and applies to every
text frame you send on that connection. There is no per-frame switch. To mix
romanized and non-romanized text, open a connection for each.
The stream has no equivalent of the X-Romanized-Applied header.
Errors
Send true or false. The server refuses strings and numbers, including
"true" and 1. Over HTTP null counts as false, but the WebSocket refuses
it.
| Endpoint | What you get |
|---|---|
POST /tts/bytes |
400 with {"error": "field 'romanized' must be true or false"} |
| WebSocket | An error frame with code: "protocol_violation" and the message romanized must be true or false, then close 1008. The session never opens. |