Dhvni turns text into natural speech and recordings into accurate text for Indian users, including the sentences that switch languages halfway. Streaming audio starts in about 0.3 seconds.
आपका refund approve हो गया है, तीन से पाँच working days में amount account में आ जाएगा।
Arjun, support
0:00 / 0:06
These demo voices were designed by our model from a written description. None of them belongs to a real person.
Most speech engines expect one language per sentence. Give them Hindi with an English word in the middle and you get the wrong accent, or a transcript in the wrong script.
Dhvni is set up for the way people here actually talk: Hindi in Devanagari and English terms in English letters, in the same sentence, in both directions.
Reads naturally, with the English terms said the way a teacher says them.
Dekho beta, mitochondria cell ka powerhouse hai.
Hindi typed in English letters picks up an English accent.
question.m4a
3.2 s voice note from a student
Hint sent with the file
NCERT, mitochondria, Priya Ma'am
Transcript
सर, mitochondria को powerhouse क्यों कहते हैं?
{
"text": "सर, mitochondria को powerhouse क्यों कहते हैं?",
"model": "dhvni-transcribe-1",
"duration_seconds": 3.2
}Give your app a voice, and let it understand the people talking to it. Both run on the same account, the same voices and the same usage report.
dhvni-voice-1Send text, get audio back as a file or as a stream that starts in about 0.3 seconds.
dhvni-transcribe-1Upload a recording, get back text written the way Indians actually type it.
Your users switch between Hindi and English in a single breath. Dhvni keeps up, in both directions.
Your teacher, your founder or your brand voice, from 10 to 60 seconds of clean audio. We set up every cloned voice ourselves, and only after the person it belongs to has agreed: they read a short phrase in their own voice, or you share a signed release that we record against the voice.
Consent recording
Priya Sharma for Sunrise Academy
“I, Priya Sharma, am the owner of this voice. I consent to Sunrise Academy creating and using a synthetic copy of my voice. My verification code is mango, rocket, violin.”
Lessons, doubt answers and revision notes read out in a teacher's own cloned voice.
आज हम photosynthesis पढ़ेंगे, पहले chlorophyll समझते हैं।
Spoken replies for IVR and in app help, and transcripts of what callers asked.
आपकी complaint register हो गई है, ticket number SMS कर दिया है।
Delivery updates, OTP calls and EMI reminders that sound like a person, not a robot.
आपका parcel out for delivery है, rider बीस मिनट में पहुँचेगा।
Bring your own LLM. Stream its words into Dhvni and play the audio back as it arrives.
जी बिल्कुल, मैं आपकी booking कल सुबह दस बजे के लिए confirm कर देता हूँ।
Plain HTTPS and WebSockets, JSON in and audio out. If your code can call a REST API, it can talk to Dhvni in an afternoon.
/v1/ttsSpeak text as a file or a stream
/v1/tts/streamSpeak text while it is being written
/v1/sttTranscribe a recording
/v1/voicesList the voices on your account
/v1/usageRead the characters and seconds you used
Speak a line and save it as mp3
curl https://api.dhvni.in/v1/tts \
-H "Authorization: Bearer $DHVNI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"voice_id": "voice_7f3a91c2d4e85b60",
"text": "नमस्ते, आज हम photosynthesis पढ़ेंगे।",
"format": "mp3"
}' \
-o lesson.mp3Step 1
Tell us what you are building. We create your account and send an API key, shown to you once.
Step 2
Send clean recordings and the speaker's consent. We train the voice for a one time ₹5,999 and add it to your account.
Step 3
List your voices, send text or audio, and read your usage from the same API.
Pay as you go with no monthly fee. Monthly plans from ₹9,999 lower the speech rate, down to ₹1.70 per 1,000 characters on Growth. Prices exclude GST.
Something else on your mind? Write to support@dhvni.in.
Hindi, English and Hinglish, including sentences that switch between the two. For speech, write Hindi in Devanagari and keep English words in English letters, for example “देखो बेटा, mitochondria cell का powerhouse है।” That reads far more naturally than Hindi typed in English letters.
With streaming turned on, the first audio arrives in about 0.3 seconds and the rest keeps coming as it is generated. Without streaming, you get the complete file as soon as it has rendered.
No. We clone a voice only when the person it belongs to has agreed, either by reading a consent phrase in their own voice that we verify, or through a signed release that we record against the voice. Our team sets up every voice in this first version.
Yes. Connect to the streaming endpoint, send text as your model writes it, and each sentence is spoken as soon as it is complete. You can flush a half sentence or cancel playback at any point.
Clean speech from one person, with no music or other voices, plus their consent. 10 to 60 seconds is enough to start; 20 to 30 minutes, for example a few recorded lessons, gives the closest match. A quiet room and a phone held close is usually enough.
In rupees, by usage. On pay as you go, speech is ₹2.00 per 1,000 characters and transcription ₹0.49 a minute, with no monthly fee. Monthly plans from ₹9,999 lower the speech rate. Training a cloned voice is a one time ₹5,999. Prices exclude GST.
120 requests a minute, 4 speech and 8 transcription requests at once, 5,000 characters of text and 25 MB of audio per request. If you hit a limit, the API tells you how many seconds to wait. Write to us if you need more.
We do not store the audio you send for transcription. Reference audio for a voice is deleted when you delete the voice; consent records are kept for audit. Our privacy policy lists every provider that processes data on our behalf.
Tell us what you are building and we will set up your account. You get: