A voice that talks, not one that presents
The voice for the agent you are building. Most speech models read like narration; Soma talks the way people actually talk, including Hinglish, written in either script.
achha dekho, itni jaldi decide mat karo na, ek week aur socho.अच्छा देखो, इतनी जल्दी decide मत करो ना, एक week और सोचो।
One line, generated by Soma. Switch the script to read the same input written the other way.
Most models read it. This one says it
Read to a room.
The register of narration, podcasts, explainer voiceover. Projected, evenly paced, built to be listened to rather than answered.
Said to you.
It stops, restarts, leans on a word, and leaves air where a person would breathe. It keeps the hmm and the haan that a person actually makes, instead of scrubbing them out.
Write it in either script
haan ji bataiye, aapke order mein kya dikkat aa rahi hai?हाँ जी बताइए, आपके order में क्या दिक्कत आ रही है?
aapki appointment kal shaam paanch baje hai, confirm kar dein na?आपकी appointment कल शाम पाँच बजे है, confirm कर दें ना?
achha ek kaam kijiye, app mein jaake status check kar lijiye.अच्छा एक काम कीजिए, app में जाके status check कर लीजिए।
What it is like to work with
- One voice, all the way through
- The same person from the first word to the last, and the same person again on the next take.
- Direction you can hear
- Say how it should sound alongside the text, in plain language. The same voice can be pointed at conversation, at narration, or at recitation.
- Pacing that reads the sentence
- Punctuation does what it looks like it should. A comma breathes, a full stop resets, and the intonation follows the sense of the line instead of a metronome.
- A voice from ten seconds
- Clone one from a short reference clip, with nothing to train and nothing to wait for. Or design one that belongs to no real person, and move it along pitch and brightness until it sits where you want it.
- English that sounds Indian
- English words inside a Hindi sentence are spoken in Indian-accented English, not dropped in from somewhere else.
- Holds to the last sentence
- A long passage runs to the end of the text, with the energy of the closing line matching the opening one.
- Keeps up with a real conversation
- Audio streams as it is generated, so your agent starts speaking without waiting on its own last word. Interrupt it and it stops where it is.
- Says the word you meant
- Hindi is full of pairs a single vowel separates, where the wrong one changes the sentence. Soma learns those from how people actually say them rather than from spelling.
Also speaks
Hinglish (romanised), Hindi, Indian-accented English, Sanskrit
Hear it on your own words
soma-tts-1 is in developer preview. It is early, and changing quickly.