Skip to content
ByteKnot Labs

AI Lab

Open models for Indian languages, small enough to run on a CPU.

The AI Lab is where ByteKnot Labs researches, trains and publishes its own models. We start from how people in India speak and write, keep each model small enough for ordinary hardware, and publish the weights, the training records and the scores, including where we lose.

Meet AksharSTT Try it in your browser

Models

What the lab has published.

Each model ships with its weights, its training records and its scores on public test sets.

  • Released v1
  • Open weights
  • Apache 2.0
  • Runs on a CPU

Speech-to-text model

AksharSTT

Hindi, English and Hinglish speech to text that runs offline on a CPU.

model file
29.5 MB
real time on a CPU
96×
mean word error rate
22.28%

Next in the lab

  • Nine more Indian languages

    Marathi, Bengali, Tamil, Telugu, Gujarati, Kannada, Malayalam, Punjabi and Odia for AksharSTT, one at a time, each shipping with its scores on public test sets.

  • A lite build of v1

    The v1 recipe in a smaller file, after akshar lite v0 showed that a 20.4 MB model is usable.

  • Better Indian-accented English

    More accented English in training, to close the gap on Svarah, AksharSTT's weakest test set.

How the lab works

Small on purpose. Measured in public.

Four rules every model we publish follows, so you can check our claims before you rely on them.

  • 01

    Small enough for ordinary hardware

    Models that run offline on the CPU people already have, fast enough to feel instant. Size and speed are measured and published alongside accuracy.

  • 02

    Scores in public, losses included

    Every number comes from one harness on full public test sets, beside a named open model, and the sets where we lose stay in the table.

  • 03

    Open weights, open records

    Weights under Apache 2.0, the full-precision checkpoint, the training records and the code, so anyone can check a result or train it further.

  • 04

    Data we can name

    Training data comes from open datasets with licences that allow commercial use, and clips that match a test clip are removed before training.

AI at ByteKnot Labs

From the lab to your product.

Client projects run on the same method: an evaluation set before the first prompt, the smallest model that measures the same, and code and weights you own.

All services
01assistants · copilots · AI-native CRM · AI integration

AI products and CRMs

Assistants, copilots and AI-native CRMs that work from your data, and AI added to the systems you already run.

Explore
02agents · RAG · fine-tuning · private LLMs

Agents, RAG and LLMs

Agents, retrieval systems and private or fine-tuned models, tested on your real cases and deployed on your infrastructure.

Explore

Guides for AI decisions

RAGvsFine-tuning

RAG vs fine-tuning: which does your AI product need?

Read the comparison
Custom AI assistantvsChatGPT Enterprise or Microsoft Copilot

Custom AI assistant vs ChatGPT Enterprise and Copilot

Read the comparison
AI-native CRMvsTraditional CRM

AI-native CRM vs traditional CRM: what actually changes

Read the comparison
AI terms in plain English

FAQ

AI Lab questions

Last reviewed

What is the ByteKnot Labs AI Lab?

The AI Lab is the research side of ByteKnot Labs. It trains and publishes small, open models for Indian languages that run on ordinary hardware, starting with AksharSTT, a speech-to-text model for Hindi, English and Hinglish.

What has the AI Lab released?

AksharSTT v1, released on 7 November 2026: speech to text for Hindi, English and Hinglish that runs offline on a CPU, in one 29.5 MB file under Apache 2.0. Its earlier versions, akshar v0 and akshar lite v0, are scored on the same test sets.

Are the AI Lab's models open source?

Yes. AksharSTT's weights and code are released under Apache 2.0, and the full-precision weights, training records and evaluation harness are public. You can use it commercially, check every score and train it further.

Why does the AI Lab build small models?

Because most machines have a CPU and no GPU, and because audio often cannot leave the device. A model that fits in 30 MB and runs offline has no per-request fee, keeps data where it is and works without a connection.

Can ByteKnot Labs build a model like this for us?

Yes. Custom models are part of our AI work: we build the evaluation set from your data first, train or adapt the smallest model that meets it, and deploy it in your own cloud account or on your devices. You keep the weights.

Taking on new projects

Need a model that runs where your data is?

A short call, a written brief, and a straight answer on scope, price and timeline before any work starts. NDA first if you need one.

Start a project [email protected]