Tamil, Telugu, Kannada, Malayalam TTS API: language codes on Sume

Beyond Hindi, Sume TTS can be pointed at eight more Indian languages by setting language to bn (Bengali), ta (Tamil), te (Telugu), kn (Kannada), ml (Malayalam), mr (Marathi), gu (Gujarati) or pa (Punjabi). Cartesia lists each of them on Sonic 3.6, Sonic 3.5 and Sonic 3, and Sume forwards the field unchanged. Odia (or) and Urdu (ur) are on Sonic 3.6 only; they have their own post.

Only Hindi is in Sume's own voice-library list. The other eight can be synthesized with a voice UUID you hold, but you cannot create and save a voice in them through Sume's voice tool.

Which models carry which language?

Cartesia's pages give the same answer for all nine codes, so the table is mostly a coverage check. The final column is the part that differs by product: it shows whether you can save a voice in that language through Sume.

Does the Hinglish support apply to these languages?

No. Cartesia's August 2026 Sonic 3.6 notes describe expanded support for Hindi written in Latin script and a normalization field for pairing a spoken language with a different number and date convention. That is described for Hindi and Hinglish; the page makes no equivalent promise for romanized Tamil, Telugu or the rest, so send those in their native script.

Sume has no normalization field. Its TTS body takes transcript, language, output_format, pronunciation_dict_id, generation_config, timestamps and segmentation. If a number must be read a particular way, write it out in words in the transcript.

What do the request limits mean for Indic scripts?

Sume counts transcript characters, up to 20,000 per request, at $0.0475 per 1,000. Indic scripts use combining marks, so a short-looking line can carry more characters than you expect; count the string you send. Each job is also capped at 1,200 seconds of audio, and a longer script fails with tts_duration_exceeded, so split a long Tamil chapter by section and join the WAV takes with Sume's audio timeline (up to 20 parts, gapless).

• Use output_format.container: "wav" for takes you will join.
• Use generation_config.speed between 0.6 and 1.5 to fit a slot.
• Pass timestamps.words: true for caption timing.

Request example

Tamil, with a Tamil-capable voice UUID in VOICE_ID:

import os
import uuid

import requests

r = requests.post(
    "https://api.sume.com/v1/tts-1.0/generate",
    headers={
        "x-api-key": os.environ["SUME_API_KEY"],
        "Idempotency-Key": str(uuid.uuid4()),
    },
    json={
        "transcript": "எங்கள் வாராந்திர செய்திக்கு வரவேற்கிறோம்.",
        "language": "ta",
        "voice": {"mode": "id", "id": os.environ["VOICE_ID"]},
        "mode": "async",
    },
    timeout=30,
)
r.raise_for_status()
print(r.json())

How should I test before a long script?

Queue a short sample of around 200 characters, which costs $0.0095 at Sume's rate, and listen for conjuncts, numerals and English loanwords. Indic scripts mix local words with English terms often, and the sample shows how the voice handles that mix. If it does not read a term correctly, spell it as it should sound. Keep the same voice id across the whole series so every chapter sounds alike, and set language identically on every request.