Live product · No signup

The whole product, live in your browser

Hear our fine-tuned voice model, watch an agent assemble itself, and talk live to a real agent — no slides, no forms.

Fine-tuned TTS

A voice model fine-tuned for your language

Most voice models treat Nepali as an afterthought. Ours are fine-tuned on human-labeled regional audio grammar, cultural nuance, and domain vocabulary all come through. The same Nepali passage, before and after.

Nepali sample

आजका प्रमुख समाचारअनुसार देशभर मनसुनी वर्षा सक्रिय रहँदा केही क्षेत्रमा दैनिक जनजीवन प्रभावित भएको छ। यसैबीच, कृषि तथा पूर्वाधार विकासका लागि करिब रु. २५ करोड बराबरका नयाँ कार्यक्रम सञ्चालनमा ल्याइएको जनाइएको छ। भारततर्फ नेपालको विद्युत् निर्यात ६० मेगावाट पुगेको छ भने पर्यटन क्षेत्रमा यस वर्ष १२ लाखभन्दा बढी विदेशी पर्यटक भित्र्याउने लक्ष्यअनुसार आगमन क्रमशः बढिरहेको बताइएको छ। साथै, डिजिटल सेवा तथा अनलाइन कारोबारको विस्तारसँगै विभिन्न सरकारी र निजी सेवाहरू थप सहज रूपमा उपलब्ध हुन थालेका छन्।

Before fine-tuning

An open multilingual base checkpoint, used as shipped. Nepali is one of many languages it was never specialised for.

Male
Female
After fine-tuning

Fine-tuned on human-labeled Nepali audio the same speaker identity, now actually reading Devanagari.

Male
Female
The model family

Every model, playable

The full set of fine-tuned speech models behind that difference. Pick one and hear the voices it ships with.

Model
English script

According to today's top headlines, global A.I startups attracted more than 2 billion dollars in fresh investments this month, while a leading cloud provider announced plans to invest 500 million dollars in expanding its data center network. Meanwhile,smartphone shipments increased by 9% year over year as consumer demand continued to recover. Additionally, several companies unveiled new AI-powered productivity tools aimed at businesses and students.

Forms

Let it take a booking, not just a message

Agents can surface a real multi-step form mid-conversation. Pick a template on the left and the live preview builds itself this is the actual Forms tab from the product.

Forms

(1)
New form
Reservation3 steps

Form Settings

Title
Heading shown at the top of the form
Reservation
Subtitle
Short context line under the heading
Reserve your spot
Method
HTTP verb
POSTPUTPATCH
Submit URL
Endpoint that receives the form payload
https://your-domain.com/api/your-endpoint

Leave empty and we'll route through Oshara's default endpoint — convenient, but it counts against your usage and may incur charges. Provide your own URL to keep submissions free.

Submit label
Final-step button text
Reserve Now
Success message
Shown after a successful submit
Your reservation is confirmed! A confirmation will be sent to your email.
Field layout
How fields stack on the page
StackGrid
Density
Spacing between rows
ComfortableCompact
Label position
Where field labels render
TopInline
Add step

Guest Details

Optional subtitle

Fields4Add field
Width
FullHalf
Required field
Live PreviewUpdates as you editStep 1 / 3

Reservation

Reserve your spot

Step 1 of 3 · Guest Details

First name*
Last name*
Email address*
Phone number
Live agent

Talk to a real Oshara agent, live

Everything above, assembled and deployed. The builder on the left is what the agent on the right runs on — tap the mic and talk to it. No login, no setup.

The honest part

A great voice is still only one voice

Fine-tuning fixes how a model speaks a language. It doesn't fix the fact that a single checkpoint is one fixed persona — and real conversations need more than one.

New product · In development

PersonaPlex

Coming soon

One fine-tuned base, many faces on top of it. PersonaPlex makes persona, voice, and tone something a live agent sets from a prompt and a short voice sample — not another checkpoint to train — and it holds a real, full-duplex conversation while it does. One deployment, ready to meet every caller the way that caller needs to be met.

The limit todayWith PersonaPlex

One voice for every caller

A single checkpoint delivers one way. The warmth that settles an anxious first-timer is the wrong read for a buyer who wants the answer in ten seconds.

A persona for each caller

Role and tone are set by a prompt, not baked into a checkpoint — so one deployment can answer each caller in the register that fits them.

It waits for you to finish

Turn-based agents listen, then speak, like a walkie-talkie. Talk while it's talking and it either speaks over you or stalls.

Full-duplex conversation

It listens and speaks at once — cut in and it stops cleanly, hears a quick “mm-hm” without losing its place, and picks the thread back up.

Code-switching breaks the voice

A Nepali call slips into English mid-sentence. Hand off between two models to cover it and the speaker audibly becomes someone else.

Both languages, one speaker

Nepali and English come from the same voice, so switching language never switches who's talking.

It can't read the room

The same flat delivery for a complaint, a sale, and a booking confirmation — tone stays fixed no matter what the moment needs.

Tone that tracks the moment

Register, pace, and warmth shift with context — patient on a complaint, brisk on a booking, reassuring on a refund.

A model per voice doesn't scale

Every new brand voice, language, or team is another fine-tune to train, host, and keep in sync as the base moves on.

New voices without a retrain

New voices come from a short reference sample, layered on the fine-tuned base — no extra training run, no second model to host.

Multi-personaFull-duplexAdaptive voiceContext-awareOne deployment
Get early access

Ready to Build the Real One?

Everything you just saw takes an afternoon to set up for real — design the persona, feed it your knowledge, generate voice and video, and deploy to your site.