a

a

a

d

d

d

a

a

a

l

l

l

a

a

a

t

t

t

a

a

a

i

i

i

Where Law Meets Artificial Intelligence.

A standalone space at the intersection of technology, justice, and real-world AI deployment — from production models and NLP to regulation, ethics, and court infrastructure.

Legal Articles

Featured Articles

Guide

15

min read

Why Whisper Cuts off Indic Transcripts after Six Seconds

Whisper silently stops transcribing Indic-script audio after about six seconds, because Indic text is token-heavy and Whisper caps each segment's output at 224 tokens. This post explains why it happens, shows it is a tokenizer-plus-decoder limit rather than a training problem, and covers the fixes — VAD chunking, a one-line cap change, or a better tokenizer — with before/after numbers.

Guide

15

min read

Building a Reliable Voice Transcription Pipeline for Indian Courtrooms (Part 1)

Indian courts move at the speed of a judge's handwriting. Adalat AI set out to change that with real-time voice transcription — but flaky courtroom internet made simply keeping the connection alive its own engineering battle. Part 1 covers WebSockets, application-level keepalives, and the hard lessons of unreliable networks.

Guide

15

min read

Vividh-ASR: Diagnosing and Fixing Studio-Bias in Whisper for Indic Languages

Fine-tuning Whisper for Hindi and Malayalam revealed a blind spot in how we train ASR models — and a counter-intuitive fix that lets a smaller model outperform much larger ones

Featured Articles

Guide

15

min read

Why Whisper Cuts off Indic Transcripts after Six Seconds

Whisper silently stops transcribing Indic-script audio after about six seconds, because Indic text is token-heavy and Whisper caps each segment's output at 224 tokens. This post explains why it happens, shows it is a tokenizer-plus-decoder limit rather than a training problem, and covers the fixes — VAD chunking, a one-line cap change, or a better tokenizer — with before/after numbers.

Guide

15

min read

Building a Reliable Voice Transcription Pipeline for Indian Courtrooms (Part 1)

Indian courts move at the speed of a judge's handwriting. Adalat AI set out to change that with real-time voice transcription — but flaky courtroom internet made simply keeping the connection alive its own engineering battle. Part 1 covers WebSockets, application-level keepalives, and the hard lessons of unreliable networks.

Guide

15

min read

Vividh-ASR: Diagnosing and Fixing Studio-Bias in Whisper for Indic Languages

Fine-tuning Whisper for Hindi and Malayalam revealed a blind spot in how we train ASR models — and a counter-intuitive fix that lets a smaller model outperform much larger ones

Featured Articles

Guide

15

min read

Why Whisper Cuts off Indic Transcripts after Six Seconds

Whisper silently stops transcribing Indic-script audio after about six seconds, because Indic text is token-heavy and Whisper caps each segment's output at 224 tokens. This post explains why it happens, shows it is a tokenizer-plus-decoder limit rather than a training problem, and covers the fixes — VAD chunking, a one-line cap change, or a better tokenizer — with before/after numbers.

Guide

15

min read

Building a Reliable Voice Transcription Pipeline for Indian Courtrooms (Part 1)

Indian courts move at the speed of a judge's handwriting. Adalat AI set out to change that with real-time voice transcription — but flaky courtroom internet made simply keeping the connection alive its own engineering battle. Part 1 covers WebSockets, application-level keepalives, and the hard lessons of unreliable networks.

Guide

15

min read

Vividh-ASR: Diagnosing and Fixing Studio-Bias in Whisper for Indic Languages

Fine-tuning Whisper for Hindi and Malayalam revealed a blind spot in how we train ASR models — and a counter-intuitive fix that lets a smaller model outperform much larger ones