Designing the front door for India’s speech reasoning lab
1 of 4
AI research labs backed by the Indian govt
7 weeks
Kickoff to public launch
8 days
from first commit to production
“Thanks folks for your tireless support in getting our website launch ready. Appreciate all the help and late night pushes in getting this across the line!”
Suvrat BhooshanCTO at Chariot- Client
- Chariot
- Industry
- Frontier AI research lab
- Services
- Website
Development
Model
Research
Team
Careers
News
Try Chariot TTS
A frontier AI research lab from
India, selected by the Government of
India and the IndiaAI Mission to
build foundational AI models.
The next reasoning
engine will speak.
Text had its GPT moment. Speech is still
fragmented into transcription, synthesis, music,
dubbing, and voice agents. Chariot is building
native speech reasoning models, systems that
understand not only words, but what is being
said, meant, and needed next.
Try Chariot TTS
Chariot is a frontier AI research lab, one of four teams selected under the IndiaAI Mission to build India’s foundation models.
The people behind it are researchers from DeepMind, Meta AI Research, Stanford and IIT, and the co-founders had already presented the work to the Prime Minister at his residence. What they hadn’t done yet was show it to everyone else: their speech reasoning models were about to go public for the first time.
We led brand direction, designed the full landing page, and built the production site, including the illustration and motion system that carries the story. Chariot-1 TTS went from private enterprise access to a public research preview after our engagement.
Generate speech
Audio is a modality, not a domain.
Coding, math, and law are domains that live inside text. Speech is not one of them — it carries meaning in timing, rhythm, emphasis, silence, accent, and social intent. Signal text never sees.
Take 500 mg twice daily for 5 days; refill Rx #88214.
Send Dr. Rao the ₹2.5L invoice, OTP 4021, before 3:30 pm.
Flight AI-302 departs T2 at 06:45, gate B12.
Your OPD slot on 12/03 is with Dr. S. Iyer, room 4B.
Paid ₹1,09,999 via UPI to BLR Motors on 2 Jan.
“Send Doctor Rao the two point five lakh rupees invoice. OTP, four zero two one. Before three thirty pm.”
Meera
Take 500 mg twice daily for 5 days; refill Rx #88214.
Send Dr. Rao the ₹2.5L invoice, OTP 4021, before 3:30 pm.
Flight AI-302 departs T2 at 06:45, gate B12.
Your OPD slot on 12/03 is with Dr. S. Iyer, room 4B.
“Send Doctor Rao the two point five lakh rupees invoice. OTP, four zero two one. Before three thirty pm.”
Meera
Chariot builds models that reason about speech. That is hard to explain and hard to show, and they were about to put it in front of the public for the first time. The website had to do the explaining.
So we gave every section its own illustration and animation. Each one shows what the model does before the text explains it. A first-time visitor can follow the page without reading all of it.
01
Generate
Natural speech from text, with context-aware delivery.
02
Understand
Speech models that hear more than words: prosody, emphasis, uncertainty, interruptions, emotion.
Chariot TTS
Your text
Dr. Mehta moved the meeting from 3:05 p.m. to 4:30. Please tell Riya it’s urgent, but don’t make it sound alarming.
10k Credits remaining
//
115/2000
Voices
0:01 / 0:20
Write naturally.
Chariot understands how it should be spoken.
Chariot TTS
Your text
Dr. Mehta moved the meeting from 3:05 p.m. to 4:30. Please tell Riya it’s urgent, but don’t make it sound alarming.
Voices
10k Credits remaining
//
115/2000
0:01 / 0:20
03
Converse
Low-latency speech-to-speech interaction with memory, turn-taking, and task understanding.
04
Reason and act
Models that can edit, role-play, teach, summarize, rewrite, and complete tasks natively in speech.
Preference ELO
Blind pairwise human
ratings across four systems
- 38
- Raters
- 57,892
- Samples
- 8
- Prompt categories
Latency
142 ms
Time-to-first-audio, median (p50).
Intelligibility
98.1%
1.9% WER Lower WER is better
Normalization robustness
97.1%
MOS spread across all three languages
Language consistency
Δ 0.06
MOS spread across all three languages
Preference ELO
1264
Blind pairwise human ratings across four systems
Latency
142 ms
Time-to-first-audio, median (p50)
Intelligibility
98.1%
1.9% WER Lower WER is better
Normalization robustness
97.1%
MOS spread across all three languages
Language consistency
Δ 0.06
MOS spread across all three languages
Chariot launched seven weeks after kickoff, with a site where every section shows what the model does before the text explains it.
We took the build from first commit to production in eight days. After our engagement, Chariot-1 TTS moved from private enterprise access to a public research preview, with the demo on the site as the way in.
