Imla — Offline Urdu Transcription

Watch

The 90-second tour — pipeline included

Demo audio is a fictional talk, synthesized and degraded to sound far-field. The transcripts, levels and timings shown are Imla’s real output on that file.

The Problem

Real Pakistani speech, recorded in real rooms

Professionals in Pakistan switch to English mid-sentence, and most recordings are a phone on a table, far from the speaker or picking up a PA system. Cloud transcription stumbles on both — and it means uploading every lecture, meeting and interview to someone else’s server.

What I Built

Drag in files, get documents back — all on the Mac

Imla is a native desktop app. A one-time setup installs ffmpeg, builds whisper.cpp with Apple Metal acceleration and downloads the Whisper Large v3 model; after that it never touches the network. Drop in up to ten audio or video files, reorder the queue, and press Start. Each file gets its input level shown up front, live progress with a time estimate, and finishes as a .docx transcript and a .srt subtitle file. A vocabulary list biases recognition towards names, places and medical terms, and Urdu, English or mixed modes are available, with an optional second pass that saves a native-script archive copy.

Technical Highlights

A pipeline hardened for far-field recordings

See a bad recording before its transcript

Duration and mean level are measured on the original file before any clean-up; anything quieter than −20 dBFS is flagged as far-field in the queue.

Audio conditioning

An 80 Hz high-pass, FFT denoise and loudness normalisation to −16 LUFS. On the demo file it lifted a −48 dBFS recording to −17 dBFS in under a second — and fixed a misheard final word.

Silence-aware chunking

Files over five minutes are cut into roughly four-minute chunks at detected silences, transcribed one by one, then stitched back with corrected timestamps.

A guard against runaway loops

Repetition and compression-ratio checks catch the confident decode loops that whisper.cpp’s own checks miss, and retry that chunk once at a higher temperature.

Mixed-language mode

English spoken mid-sentence comes through as English rather than being forced into Urdu script — the way code-switched speech is actually read.

Built for a v2

An on-device LLM stage for translation and content drafts is already wired into the pipeline and disabled; a failing post-processor can never break a transcript.

Stack

Under the hood

Desktop shell	Tauri with a Rust backend — setup, ffmpeg extraction, whisper.cpp invocation, output writers
Frontend	React + TypeScript with Zustand driving the sequential batch queue
Speech model	Whisper Large v3 via whisper.cpp, built locally with Apple Metal acceleration
Outputs	Right-to-left-aware .docx transcripts and timestamped .srt subtitles, with resume/skip of finished files

Status

Where it stands

Shipped as an installable Mac app, with the far-field pipeline hardening added since. Version 2 will switch on the on-device language-model stage for translation and drafting.