Languages

10 languages a caller can hold a conversation in. 11 understood by the transcriber we ship, 15 scheduled languages plus Indian English speakable in a live call, 26+ detectable from audio.

Two-way is the number that mattersUnderstanding a language and being able to reply in it are different capabilities, and only the overlap lets someone have a conversation. That overlap is 10. The wider "understood" figures elsewhere in this documentation counted models that need an explicit language code, want 16kHz audio, or have no streaming path — none of which a live 8kHz call can use. Check the routing table below before promising a language to anyone.

Live-call routing table

Which streaming TTS model can speak each scheduled language. Anything showing none can be transcribed but not spoken in real time.

Language	Code	Understand	Detect	Speak
Assamese অসমীয়া	`as`	✓	✓	male & female
Bengali বাংলা	`bn`	✓	✓	male & female
Dogri डोगरी	`doi`	✓	✓	male & female
Gujarati ગુજરાતી	`gu`	✓	✓	male & female
Hindi हिन्दी	`hi`	✓	✓	male & female
Kashmiri کٲشُر	`ks`	✓	✓	not yet
Konkani कोंकणी	`kok`	✓	✓	not yet
Maithili मैथिली	`mai`	✓	✓	male & female
Marathi मराठी	`mr`	✓	✓	male & female
Nepali नेपाली	`ne`	✓	✓	male & female
Odia ଓଡ଼ିଆ	`or`	✓	✓	not yet
Punjabi ਪੰਜਾਬੀ	`pa`	✓	✓	male & female
Sanskrit संस्कृतम्	`sa`	✓	✓	male & female
Sindhi سنڌي	`sd`	✓	✓	not yet
Urdu اردو	`ur`	✓	✓	not yet
Kannada ಕನ್ನಡ	`kn`	✓	✓	male & female
Malayalam മലയാളം	`ml`	✓	✓	male & female
Tamil தமிழ்	`ta`	✓	✓	male & female
Telugu తెలుగు	`te`	✓	✓	male & female
Bodo बड़ो	`brx`	✓	✓	male & female
Manipuri ꯃꯤꯇꯩ ꯂꯣꯟ	`mni`	✓	✓	not yet
Santali ᱥᱟᱱᱛᱟᱲᱤ	`sat`	✓	✓	not yet

Cannot be spoken at allKashmiri, Konkani, Odia, Sindhi, Urdu, Manipuri, Santali. These can be transcribed and translated, but no deployed text-to-speech model has a voice for them, so they cannot be spoken — not mid-call and not to a file. To reply, translate into a language that does have a voice, or fall back to text.

What an agent can converse in: 10 languages, 20 preset voices, plus your own cloned onesMale and female for each of Hindi, Bengali, Marathi, Telugu, Kannada, Tamil, Malayalam, Gujarati, Punjabi and Indian English. There are another 18 voices for the nine languages we can speak but cannot transcribe — Assamese, Nepali, Sanskrit, Maithili, Bodo, Dogri, Bhojpuri, Magahi and Chhattisgarhi. Those 38 were all synthesised and quality-checked, and all 38 produced usable audio, but only the 20 above are offered to an agent, because the other nine cannot hear a reply.

Every count on this page is a language we will actually speak todayNothing here is counted on the strength of a demo or a roadmap. If a language appears, an agent can hold a call in it now, and you can hear both of its voices before you choose.

Beyond the scheduled languages

We also speak three widely-used non-scheduled languages that most vendors skip entirely:

Language	Code	Native
Bhojpuri	`bho`	भोजपुरी
Magahi	`mag`	मगही
Chhattisgarhi	`hne`	छत्तीसगढ़ी

Quality is not flat

Coverage and quality are different things. Nobody — including the large commercial vendors — has good training data for every Indian language.

Tier	Languages	What to expect
strong	Hindi, Tamil, Telugu, Bengali, Marathi	Ship these. Plenty of training data.
usable	Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese	Good enough, test on your own audio first.
weak	Bodo, Santali, Dogri, Manipuri, Sanskrit, Kashmiri, Sindhi, Konkani	Listed as supported, but the data does not exist to make them good. Do not promise equal quality.

Code-mixing and Hinglish

Most real Indian phone calls are not one language. They switch between a regional language and English inside a single sentence. That is the normal case, not an edge case.

Synthesis handles it: a sentence that mixes Hindi and English is spoken as one sentence, not as two accents stitched together. Transcription is the weaker side, and we would rather say so than let you discover it — a heavily mixed utterance is transcribed less reliably than a monolingual one. Closing that gap properly means training on real call recordings from traffic like yours.

Mid-call language switching

Supported, on by default, and the voice stays the same through it. A caller can open in Telugu, move to Hindi and finish in Tamil, and hear one person the whole way.

That last part took work and is worth knowing aboutVoices are defined per language, so the naive behaviour is that changing language changes the speaker and the caller feels handed to a colleague mid-sentence. lock_voice is on by default and holds the identity across the switch. Measured on a live Telugu → Hindi → Tamil call, similarity to the opening voice was 0.818 with it on and 0.437 with it off — where two takes of one real person score 0.747 and two different people score 0.213.

Telugu to Tamil is a reliable switch. Hindi to Urdu is not detectable by audio. See the accuracy table →

Cheapest accuracy win availableSet the language from your CRM record at call start. You know who you are dialling. That is right about 95% of the time, costs nothing, and beats detection. Use detection as the safety net.

All 26 language codes

Code	Language	Native	Family	Script	Scheduled
`as`	Assamese	অসমীয়া	Indo-Aryan	Bengali-Assamese	✓
`bn`	Bengali	বাংলা	Indo-Aryan	Bengali-Assamese	✓
`doi`	Dogri	डोगरी	Indo-Aryan	Devanagari	✓
`gu`	Gujarati	ગુજરાતી	Indo-Aryan	Gujarati	✓
`hi`	Hindi	हिन्दी	Indo-Aryan	Devanagari	✓
`ks`	Kashmiri	کٲشُر	Indo-Aryan	Perso-Arabic	✓
`kok`	Konkani	कोंकणी	Indo-Aryan	Devanagari	✓
`mai`	Maithili	मैथिली	Indo-Aryan	Devanagari	✓
`mr`	Marathi	मराठी	Indo-Aryan	Devanagari	✓
`ne`	Nepali	नेपाली	Indo-Aryan	Devanagari	✓
`or`	Odia	ଓଡ଼ିଆ	Indo-Aryan	Odia	✓
`pa`	Punjabi	ਪੰਜਾਬੀ	Indo-Aryan	Gurmukhi	✓
`sa`	Sanskrit	संस्कृतम्	Indo-Aryan	Devanagari	✓
`sd`	Sindhi	سنڌي	Indo-Aryan	Perso-Arabic	✓
`ur`	Urdu	اردو	Indo-Aryan	Perso-Arabic	✓
`kn`	Kannada	ಕನ್ನಡ	Dravidian	Kannada	✓
`ml`	Malayalam	മലയാളം	Dravidian	Malayalam	✓
`ta`	Tamil	தமிழ்	Dravidian	Tamil	✓
`te`	Telugu	తెలుగు	Dravidian	Telugu	✓
`brx`	Bodo	बड़ो	Sino-Tibetan	Devanagari	✓
`mni`	Manipuri	ꯃꯤꯇꯩ ꯂꯣꯟ	Sino-Tibetan	Meitei / Bengali	✓
`sat`	Santali	ᱥᱟᱱᱛᱟᱲᱤ	Munda	Ol Chiki	✓
`bho`	Bhojpuri	भोजपुरी	Indo-Aryan	Devanagari	—
`mag`	Magahi	मगही	Indo-Aryan	Devanagari	—
`hne`	Chhattisgarhi	छत्तीसगढ़ी	Indo-Aryan	Devanagari	—
`en-IN`	Indian English	Indian English	Germanic	Latin	—

Voices, tiers and measured accuracy at /docs/models. The live list of every voice is GET /v1/models, which is also what the dashboard's voice picker reads.