Updates · Podcast
Hebrish Updates
Voice notes from Daniel on building Hebrish: what's new in the models and the apps, what worked, what didn't, and what's next. Every episode is also an article you can read, with the full transcript and the sources.
-
Cleaning up phone recordings for Meorav v2.2, and speaker isolation
How audio-to-audio models turned my speech-to-text recordings into training data for Meorav v2.2, why it still sounds flat, and why speaker isolation has me rethinking dictation as a chain of small models.
-
Why I built this, 25 hours of recordings and the two models I'm keeping
A voice note from home on a sick day: why the site now says why Hebrish exists, what recording 25 hours of sentences is like, and which models I'm committed to retraining.
-
Recording on Jaffa Street: accuracy, cleanup models and more voices
Recorded on Jaffa Street in Jerusalem: where accuracy caps out, what belongs in the speech model versus post-processing, number and date formatting, disfluencies, learning from my own mistranscriptions, and synthetic voices through IPA.
-
What Hebrish is, how it's made, and the next sprint
What Hebrish is, why it exists, how the models are trained, the model lines, how the training data is recorded, and the plans for the desktop and Android apps.
Subscribe to Hebrish Updates on Spotify, or in any podcast app with the feed.