Speech recognition
A speech model of our own
Trained by us, on Indian speech, running on our own machines. Nobody here speaks one language for a whole sentence, and an engine built for clean American English hears stock mein and writes stock may. Ours is trained on the switch itself.
What the caller said
aaj stock mein red saree hai kya bhaiya
A general-purpose English engine
Are stock may read sorry hi kya via
Dhvanii
आज stock में red साड़ी है क्या भैया
Same audio, transcribed by a general-purpose engine and by Dhvanii. Pick a line to compare.
- Live calls
- Streaming
- Words arrive while the caller is still speaking.
- Recordings
- Batch
- Support calls and meetings, transcribed in bulk.
- Two scripts
- हिंदी + English
- Each language in its own script, on one line.
- Where it runs
- India
- Our cloud, or inside yours.
What we had to solve
None of these are edge cases. They are what the first minute of an ordinary Indian phone call sounds like, and each one breaks an engine built for a different country.
Two scripts in one sentence
A transcript that romanises Hindi is unreadable to a Hindi speaker, and one that transliterates English into Devanagari is unreadable to everyone. We keep each language in its own script, decided word by word.
Numbers as they are spoken
बयालीस सौis 4,200, not “42 100”. Lakhs and crores are the units people say prices in, and a transcript that cannot write them back correctly is useless for anything billing-related.
Phone-quality audio
Most benchmarks are read speech from a good microphone. Real input is a patchy mobile line with a ceiling fan overhead and a second conversation in the room. We tune for that case, not the clean one.
Already carrying production traffic
Chiku AI
A live interview assistant used across India. Every word its candidates hear is transcribed here, in real time, while they are mid-sentence.
chiku-ai.inDhvanii voice agents
The phone agents on the other half of this site. Transcription is the first and largest slice of their one-second reply.
Voice agentsSend us an hour of your audio
The only benchmark that settles anything is your own calls. We will transcribe a sample against whatever you use today and send both transcripts back side by side, no charge.
Request API access- support@dhvanii.com
- Interface
- WebSocket, REST batch
- Data residency
- India