We benchmarked every open Amharic speech model on data none of them had seen.
The 606M model beat the 963M one.
The monolingual beat the multilingual.
CER and WER disagreed on the winner.
Then we built a 538MB language model that fixes 3 more words per 100.
Try it on your own voice 👇
We just released our first batch of open Amharic speech data by @snapwre
22.7 hours · 7,405 clips · 320 speakers
CC BY 4.0, free on Hugging Face
It's a first batch. We're not claiming it's the best quality yet
we'd rather you try it and tell us what's wrong with it.
Next release: properly preprocessed audio, a real validation pipeline.
Interested? DM me.
https://t.co/2RXn4HNCnp
We just released our first batch of open Amharic speech data by @snapwre
22.7 hours · 7,405 clips · 320 speakers
CC BY 4.0, free on Hugging Face
It's a first batch. We're not claiming it's the best quality yet
we'd rather you try it and tell us what's wrong with it.
Next release: properly preprocessed audio, a real validation pipeline.
Interested? DM me.
https://t.co/2RXn4HNCnp
Shape the Future of AI & Earn Rewards!
Help AI learn to speak Amharic, Afaan Oromo, Tigrinya, Somali, and 80+ Ethiopian languages. Record, validate, and grow the dataset today! 🇪🇹
https://t.co/PdDJBtrQC1
Our FreeSWITCH "Connection Refused" errors weren't a network problem.
They were a Go race condition in an upstream library — hidden by a systemd stderr redirect that kept the panic out of journalctl.
The trail, three wrong theories, and the fix:
https://t.co/cqT2bidcAV
How we built a VoIP platform handling millions of calls a day:
Asterisk dialplan → AGI → ARI + a custom call engine → FreeSWITCH → Kamailio load-balancing a fleet of nodes.
+ the lessons: lock down your SIP edge, keep the DB off the call path, design IVRs humans don't hate.