Recording & transcription

How transcription works, and how fast it is

Parakeet TDT turns speech into private, offline-ready text after targeted pre-transcription enhancement.

Updated 2026-07-11

Your words become text on your device with NVIDIA Parakeet TDT speech recognition, so the recording does not need to be uploaded to an Obsidian Ridge Labs server for transcription.

Before Parakeet receives the audio, Echo Chamber applies a targeted, speech-focused filter designed for transcription. This is not generic normalization. The complete enhanced Echo Chamber pipeline has produced an internal observed word error rate of approximately 4.5% under tested conditions. Results vary with speakers, accents, acoustics, crosstalk, vocabulary, and source quality.

Speed depends on your device and the length of the recording, but short notes are usually ready the moment you stop, and longer recordings finish processing quickly in the background.

A finished transcript with labeled speakers and timestamps
A finished transcript, with speakers labeled and words tied to timestamps.
Models download once

The speech model downloads a single time over the internet. After that, recording and transcription keep working fully offline.

Related guides

Why a word is wrong, and how to improve accuracy
Practical ways to get cleaner transcripts and fix the words that slip through.
Edit a transcript, tap to play, and add vocabulary
Fix any word inline, jump to that moment in the audio, and teach Echo Chamber your terms.
Using Echo Chamber offline
What works without internet, and the one thing that needs it first.
Translate a transcript
Turn a finished transcript into another language, on your device.
Still need a hand?

Email support@obsidianridgelabs.com, a real person will reply.

View Echo Chamber