Chariot-1 TTS
Realistic text to speech in 15+ languages.
Pick from 50+ voices or clone your own in seconds. Every voice speaks English, German, Chinese, Hindi, Urdu and nine more Indian languages, and reads numbers, dates and prices the way a person says them.
Free to start · No card · 10,000+ people use Chariot
इस दिवाली, घर की हर ज़रूरत पर पाइए 40% तक की छूट। ऑफ़र सिर्फ़ 31 अक्टूबर तक!
Kabir[M]Native Hindi10,000+people already use Chariot
50+ voices
Text to speech that understands what you mean.
Write the line the way you would say it. The voice hears the whisper, the bad news or the good news in the words and reads it that way, without tags or settings. Press any voice to hear it.
I found the note you left in the book. It caught me by surprise, and it made my whole day. What a wonderful surprise!
Mia[F]Soft and reflective. Meditation, sleep stories.Indian, British and American accents. Every voice speaks every language. Try them with your own text
Written text
Say what you mean. The model handles how to say it.
Traditional TTS makes you do part of the model's job: expand numbers, spell out abbreviations, add pronunciation hints. Chariot is built for the text people actually write, and uses context to decide what should be spoken, and how.
You write
Order two zero four eight contains three items and ships tomorrow.
Benchmark · October 2026
Evaluated where the bar is highest.
Chariot-1 TTS and eight other commercial systems read the same sentences through their public APIs, judged by the same three speech recognisers and one language-model judge. A correct spoken form of written text never counts as an error.
Accuracy on hard examples
Turning written forms into spoken ones: dates, currencies, URLs, units. 1,029 English sentences written the way text actually arrives, judged on whether each was said correctly.
- #1Customer service96.0%
- #1Technical & scientific95.6%
Ahead of all eight other systems. Higher is better.
Naturalness
How human the audio sounds, rated clip by clip (NISQA, 1 to 5).
Tied first in English, first in Hindi. Higher is better.
By name
The top of each table. Every system was run through its public API in October 2026, with the versions shown.
| System | Read correctly | NISQA |
|---|---|---|
| Chariot Chariot-1 TTS | 94.0% | 4.91 |
| ElevenLabs v3 | 93.5% | 4.87 |
| Cartesia Sonic-3.5 | 89.9% | 4.83 |
| xAI Grok | 89.5% | 4.91 |
| System | Hindi CER | NISQA |
|---|---|---|
| Chariot Chariot-1 TTS | 7.73% | 4.974 |
| xAI Grok † | 7.99% | 4.966 |
| ElevenLabs v3 | 8.21% | 4.867 |
| Cartesia Sonic-3.5 | 8.30% | 4.823 |
Two native Hindi voices per model. † No native Hindi voice: the system's default voice read the Hindi sentences.
Hindi error rate
Character error rate on 376 native sentences plus 268 general and code-mixed sentences, the lowest of the seven systems tested.
Lower is better; the scale runs from worst to best.
Time to first audio
The median time to the first byte of audio over the WebSocket endpoint, so a voice agent starts speaking almost at once.
Languages
15+ languages. Every voice speaks all of them.
Native voices for Hindi, Urdu and nine more Indian languages, with American, British and Indian English. Any voice reads any of them, so one voice can carry a brand from one market to the next. Each Indian language has its own page with samples.
5 more Indian-accent voices (Assamese, Nepali, Sindhi) speak every language too. Hinglish, Tanglish and Hindi typed in Roman letters are read as written.
Use cases
What people make with it.
Six kinds of work, each with a clip you can play. Every one has its own page with samples, limits and prices for that job.
Voice agents and IVR
Sentence-by-sentence streaming over WebSocket, about 75 ms to first audio, so a caller never waits for the whole reply.
Thanks for calling. I found the charge you mentioned, and I can help you sort it out.
Matt · American EnglishAds and promos
One script, every language and region. Offers, dates and rupee amounts read the way a presenter says them.
This fall, get up to 40% off everything for your home. Offer ends October 31!
Ava · American EnglishYouTube and explainers
Narration for tutorials, product videos and reels, with commercial rights on every paid plan.
In this video, we will set up your account, connect your first app and send a test message, all in under five minutes.
Nathan · British EnglishAudiobooks and stories
Voices that carry a scene: a whisper stays a whisper, good news sounds like good news.
I found the note you left in the book. It caught me by surprise, and it made my whole day. What a wonderful surprise!
Mia · American EnglishE-learning
Lessons, courses and training in the languages your learners speak, with numbers and units read correctly.
Lesson three: fractions. A fraction is a part of a whole. Half a pizza is one of two equal pieces, so we write it as one over two.
Alan · American EnglishReading aloud
Articles, documents and app content read naturally for people who prefer to listen or cannot read the screen.
Good morning. Here are today’s headlines, read aloud. The new metro line opens on Monday, and light rain is expected by evening.
Susan · British EnglishVoice cloning
Your voice, in every language.
Record or upload 30 seconds of your voice. About 20 seconds later, it reads anything you type, in every language Chariot speaks.
Aditi[F]A voice you clone works the same wayHello! I can speak English too, in my own style.
Studio
Whole voiceovers, in Studio.
For more than one line, bring your script to Chariot Studio, or just the idea. It casts a voice for every speaker, scores the piece with music and sound, and lets you change it one line at a time.
API
From text to speech in a few lines.
One endpoint, one API key. Download a finished file, stream raw audio as it generates, or hold a WebSocket open for live agents.
import os, requests MEERA = "3c763516-1d51-4d0a-b328-57b5507fb20b" resp = requests.post( "https://api.chariot.in/v1/tts", headers={ "chariotai-api-key": os.environ["CHARIOT_KEY"], "Content-Type": "application/json", }, json={ "voice_id": MEERA, "text": "Namaste! Aapka order bas pahunchne wala hai.", "model_type": "chariot-multilingual-v1", }, ) resp.raise_for_status() open("chariot-tts.wav", "wb").write(resp.content)
import os, requests MEERA = "3c763516-1d51-4d0a-b328-57b5507fb20b" with requests.post( "https://api.chariot.in/v1/tts/stream", headers={ "chariotai-api-key": os.environ["CHARIOT_KEY"], "Content-Type": "application/json", }, json={ "voice_id": MEERA, "text": "Namaste! Aapka order bas pahunchne wala hai.", "model_type": "chariot-multilingual-v1", }, stream=True, ) as resp: resp.raise_for_status() with open("chariot-tts.pcm", "wb") as f: for chunk in resp.iter_content(4096): f.write(chunk) # raw 16-bit PCM, 44100 Hz mono. To WAV: # ffmpeg -f s16le -ar 44100 -ac 1 \ # -i chariot-tts.pcm chariot-tts.wav
import os, json, asyncio import websockets # pip install websockets URL = ( "wss://api.chariot.in/v1/tts/ws" "?voice_id=3c763516-1d51-4d0a-b328-57b5507fb20b" "&model_type=chariot-multilingual-v1" "&response_format=pcm" "&split_granularity=sentence" "&api_key=" + os.environ["CHARIOT_KEY"] ) async def main(): async with websockets.connect(URL) as ws: await ws.send(json.dumps({ "type": "input.text", "text": "Namaste! Aapka order aa raha hai.", })) await ws.send(json.dumps({"type": "input.done"})) pcm = bytearray() async for msg in ws: if isinstance(msg, bytes): pcm += msg elif json.loads(msg)["type"] == "session.done": break open("chariot-tts.pcm", "wb").write(pcm) asyncio.run(main())
Pricing
Simple pricing that scales with you.
Start free. Get more minutes for your money on higher plans.
Unlimited audio downloads on every paid plan.Included on all paid plans. Audio generation uses your plan credits.
Minutes and cost per minute are estimates based on 1,000 credits ≈ 1 minute of audio and using all included credits. Actual duration varies with language and speaking pace. Prices exclude applicable taxes.
Need more? Talk to us ›
Your data
Your text and audio stay in India.
TTS runs on Indian infrastructure. Inputs and generated audio are processed in India under the Digital Personal Data Protection Act. Chariot is one of four teams selected under the IndiaAI Mission to build the country’s foundation models.
FAQ
Questions, answered.
Everything a builder asks before the first request.
Text to speech (TTS) turns written text into spoken audio. Modern systems like Chariot-1 TTS use a neural model trained on speech, so the voice carries natural rhythm and emphasis instead of reading word by word. You type or send text, pick a voice, and get an audio file or a live stream back.
Yes, to start. Every account gets 10,000 free credits, about 10 minutes of audio, with every voice and language, and you can download your first 5 clips. No card needed. Paid plans add unlimited downloads, commercial rights and voice cloning.
In Chariot’s October 2026 evaluation against eight commercial systems, Chariot-1 TTS read hard written text correctly 94.0% of the time, the highest score, and rated 4.91 out of 5 for naturalness in English and 4.97 in Hindi, tied first and first.
Chariot-1 TTS speaks 15+ languages, including English, German, Chinese, Hindi, Urdu, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Odia and Punjabi. It also reads Hinglish and other mixes of an Indian language with English, and Hindi typed in Roman letters. Every voice speaks every language, including voices you clone.
Yes. Write Hindi in Devanagari or in Roman letters the way you would text it, and mix in English words either way. The voice keeps one accent across both. The same goes for Tanglish, Banglish and the other mixes.
The way a person says them. ₹2,450 is a rupee amount, 15 March is a date, 06:45 is a time, 500 mg is a dose, Dr. is Doctor and B+ is B positive. Order and flight numbers are read digit by digit, as people do. You never need to spell these out first.
First audio arrives in about 75 milliseconds (median) over the WebSocket API, and speech streams as it generates. The WebSocket endpoint speaks sentence by sentence, so a voice agent can start talking before the whole reply is written.
Yes. Stream 16-bit PCM at 44.1 kHz and resample for telephony, or hold the WebSocket endpoint open for live conversation. Support lines, order updates and reminders in Hindi, Hinglish or any supported language use the same API. See text to speech for voice agents.
Yes, on every paid plan. Record in the browser or upload 10 to 60 seconds of one speaker (WAV, FLAC, MP3 or M4A, up to 20 MB). About 30 seconds of clean speech, without music or background noise, gives the closest match. The voice is usually ready in 10 to 20 seconds. Starter includes 3 cloned voices and larger plans include more. Moving to a smaller plan never deletes a voice: voices over the new allowance pause until you upgrade.
Yes. A cloned voice speaks every language Chariot speaks, whatever language the recording was in.
Only a voice you have the right to use: your own, or that of someone who has given you permission. You confirm this when you upload, we keep a record of it, and the Acceptable Use Policy applies.
MP3 or WAV from the web app. The API returns a WAV file, or streams raw 16-bit PCM at 44.1 kHz mono, which you can convert to any format.
Each API request takes up to 1,000 characters. For a longer script, send one request per paragraph; the voice and settings stay the same across them.
Yes. Commercial usage rights are included on every paid plan, so you can publish to YouTube, ads, podcasts and products. The free plan is for trying the voices.
Yes, in Chariot Studio. Paste a script or describe the piece, and Studio casts a voice for every speaker, adds music and sound that start on the words they belong with, and lets you change one line at a time. It uses the same voices and credits as text to speech.
Send a POST request to https://api.chariot.in/v1/tts with your API key in the chariotai-api-key header and a JSON body with voice_id, text and model_type set to chariot-multilingual-v1. The response is a WAV file. The code sample above shows the full request, with streaming and WebSocket versions.
In Chariot’s October 2026 evaluation, run through both public APIs on the same sentences, Chariot-1 TTS read hard English text correctly 94.0% of the time against 93.5% for ElevenLabs v3, scored 4.91 against 4.87 for English naturalness, and had a Hindi character error rate of 7.73% against 8.21%. ElevenLabs supports more languages overall; Chariot covers 15+, with native voices for 11 Indian languages.
In the same evaluation, Sarvam Bulbul v3 read hard English text correctly 88.2% of the time with a Hindi error rate of 8.95%, and Microsoft Azure MAI-Voice-2 scored 83.1% and 10.06%, against Chariot’s 94.0% and 7.73%. Google Cloud text to speech was not part of the test. Every system was used through its public API in October 2026.
TTS inputs and generated audio are processed in India. Website and product analytics may be processed by providers described in our Privacy Policy and Cookie Notice. You can ask us not to use your content to improve our models by writing to [email protected].
You spend credits as you generate audio, so you only pay for what you use. Revising a line is cheap rather than a whole new studio booking, and all pricing is in rupees with no forex markup.
Give your words a voice today.
Join 10,000+ people making audio with Chariot. Start free with 10,000 credits, no card needed.