EMIAC
Contact sales

Products

ClarioCrux, in alphaAirNoteTimoClario C1

Solutions

WhatsApp MCP

Enterprise

AI builds for companiesWhat we have builtSecurity and trust

Company

AboutClientsNews Contact sales

Clario C1 · our first model

A speech model for meetings
that switch languages.

Whisper base, fine-tuned on real Hindi and English meeting audio. Small enough to run on a laptop, with no network, and trained for exactly one job.

Try it in Clario A model for your company
74MParameters, the size of Whisper base
41 hOf Hinglish speech in training
0Network requests to transcribe

Why we trained our own

Work does not sound like a read-aloud dataset.

General speech models learn mostly from clean, read-aloud audio in one language at a time. A real meeting has a room, a laptop microphone, people talking over each other, and sentences that start in Hindi and finish in English. That is the speech C1 was trained on.

  • Hindi and English in the same sentence
  • English work words kept as people say them
  • Meeting rooms and calls, not a studio

Readable by default

Same meeting.
Two transcripts.

Weekly sync00:00

A general speech modelDevanagari

0 English words rewritten

Clario C1Roman, by default

0 English words kept

Illustration of C1's default output, with sample sentences.

How C1 was trained

Better data,
not a bigger model.

  1. 01

    Start from Whisper base

    OpenAI's open speech model, 74 million parameters. Asked for Hindi, it writes Devanagari, English words included.

  2. 02

    Round one: Hinglish in Roman letters

    Oriserve's open fine-tune of Whisper base, released under Apache 2.0, taught it to write Hindi the way people type it and keep English words in English. We started from there.

  3. 03

    Gather everything

    For round two, hours of audio of every kind: podcasts, audiobooks, music, broadcasts, and meetings.

  4. 04

    Keep only meetings

    Anything that is not people working together is left out, however clean it sounds.

  5. 05

    Check every clip, then train round two

    Noisy, misaligned and half-silent clips go too. The 41 hours of meeting speech that pass make it a meeting specialist.

  6. 06

    Listen to a real meeting

    What comes out knows one kind of speech well: people at work, in Hindi and English at once.

Illustration. The outputs, file names and meeting are samples.

Run it

One command,
on the device.

Inside Clario, C1 runs through whisper.cpp on the Mac's GPU. Silence is cut first, then every stretch of speech comes back as a timed line, in the language it was spoken in.

zsh · weekly-sync.wavIllustration 
$ whisper-cli \
    -m models/ggml-clario-hinglish-41h.bin \
    -f weekly-sync.wav \
    --vad --vad-model models/ggml-silero-v5.1.2.bin

# C1 on the GPU through Metal · silence removed first

[00:00:00.000 --> 00:00:04.200]  Okay, aaj ka agenda simple hai.
[00:00:04.200 --> 00:00:08.900]  Friday ko deploy karna hai, QA done?
[00:00:08.900 --> 00:00:13.400]  Haan, staging pe sab green hai.
[00:00:13.400 --> 00:00:18.100]  Invoice wala bug client ne phir raise kiya.
[00:00:18.100 --> 00:00:22.600]  Main aaj evening tak fix push kar dunga.

On the device

Built to run where the meeting is.

C1 ships inside the app. There is nothing to download at first launch, and turning speech into text makes no network request at all.

whisper.cpp, with Metal

A C++ runtime built for Apple silicon, running the model on the Mac's GPU. No Python, no server, no GPU in a data centre.

Silence removed first

A small voice-activity model cuts the silence out before C1 listens, so the computer spends its time on speech.

Checked before every use

The model file's size and SHA-256 are checked before it runs. A missing or changed file is a visible error, never a quiet switch to another model.

Works offline from the first minute

Because the model is part of the app, it works right after installation, on a plane or in a room with no Wi-Fi.

Small, on purpose.

Fast on a laptop 74 million parameters run comfortably on the computer in front of you.
Private by design The audio is turned into text on the device, not sent to be transcribed.
Trained for one job Instead of every language, it learned the speech our users actually have.

Where C1 runs today

Inside our own products.

Clario

Meeting notes you can check. C1 writes the transcript on your computer.

AirNote

EMIAC's own dictation. C1 recognises the speech on Apple silicon Macs.

Your product

A speech model trained on your team's words, accents and meetings.

Questions

Does C1 need the internet?

No. C1 ships inside the app and turns speech into text on the device, so transcription works with no connection at all.

Which languages does it understand?

Hindi and English, including the very common case where one sentence uses both.

What is C1 built on?

OpenAI's Whisper base, through Oriserve's open Hinglish fine-tune of it, Whisper-Hindi2Hinglish-Swift. Both are released under the Apache 2.0 licence. We then fine-tuned it again on meeting speech.

What script does it write Hindi in?

Roman letters, by default, with English words kept in English. "Kal ki client call 11 baje shift kar do" rather than the same sentence in Devanagari.

How big is it?

74 million parameters, stored as about 141 MB of fp16 weights. It is part of the app download, not a separate one.

Where can I use it?

In Clario today, and inside EMIAC in AirNote. If you want speech recognition built into your own product, talk to us.

Can you train a model for our company?

Yes. The same process that produced C1 can be pointed at your meetings, your words and your accents. Contact sales and tell us about your audio.

A model trained on how your team talks.

Use C1 in Clario today, or have us train one for your company.

Contact sales Try it in Clario