Building AI agents and scaling Voice AI / CPaaS at @snapwre. Also building dataset.et.
Dubai
- Tweets 332
- Following 284
- Followers 565
- Likes 1,547
Still going on with no sweat! with full back up to R2 using @getdokploy.
bytkit.com is also on this server, check it out
9
279
I love the Claude in chrome integration control with the Claude code.
Now Claude can access any website scroll it when needed while other agent building your code in parallel!
Did you all try this feature yet?
2
9
443
We open-sourced Addis Scribe Streaming, realtime speech recognition for Amharic.
It beats Google Chirp 3 on both accuracy and speed:
Word error rate: 28.4% vs 32.5%
First words on screen: 1.56s vs 6.67s
Final text after you stop talking: 30ms vs 1,411ms
Open weights, and available through our API and playgrounds.
2
2
19
474
Chapi Menge retweeted
our random side project just got shown by Stripe’s Head of Design on stage at Lenny’s Summit 😭
Katie Dill (@lil_dill) actually put it in her talk
still can’t believe this lol
here’s the moment: youtu.be/GLvFTMtw4Jk?t=1070
Stripe dashboards are boring.
So I built mine as a forest.
Every tree is a real subscriber.
Big customers grow into big trees.
When someone is about to cancel, their tree turns orange.
When they churn, the tree dies.
When a payment comes in, it rains.
You can scroll back in time and watch your whole business grow.
It's not launched yet.
Should I ship it?
@tibo_maker @levelsio @robj3d3 @marclou @jackfriks
12
7
122
6,390
Chapi Menge retweeted
Now this is good news have you guys heard of Ethiopia's first natively developed #ASR ሆሄ ?
vt.tiktok.com/ZSbNM86g8/
Keep em coming
@chapimenge3
1
2
106
We opened the server that runs Hohe ASR, our Amharic speech to text model.
Three commands, no GPU, no API key, no account:
git clone github.com/snapwre/hohe-serv… docker build -t hohe-asr . && docker run -p 8080:8080 hohe-asr curl -F audio=@clip.ogg localhost:8080/transcribe
Clean machine, timed: model loaded in 6.8s, 6s of Amharic transcribed in 1.1s.
1
25
691
Hohe ASR is out: open Amharic speech to text, built in Addis Ababa.
A five second voice note transcribed in 0.39 seconds. On a CPU. No GPU anywhere.
16.1% word error on speakers it has never heard.
Try it, nothing to install: t.me/dataset_ai_bot Weights: huggingface.co/snapwre/hohe-…
6
15
80
2,522
2/ Why it is that fast: it is a CTC model. One pass over the audio, no word by word generation. Roughly 5x faster than a Whisper of similar accuracy, and it does not hallucinate sentences out of silence.
3/ Measured on four ordinary CPU cores with int8: 5 seconds of audio in 0.39s, 30 seconds in 7.75s. An hour of recording in under ten minutes, on a machine that rents for about nine cents an hour.
4/ Accuracy holds across read speech, five regional dialects and the public FLEURS benchmark, all around 16% word error. Not tuned to the shape of our own recordings.
5/ Where it fails: spontaneous conversation, near 47% word error when people talk over each other. Numbers come out spelled as words. No punctuation. Published in the model card next to the good numbers.
6/ Full write up: dataset.et/blog/hohe-asr-amh…
1
3
193
So much appreciated! Dataset.et is working to improve Ethiopian language ai landscape
Very trendy in Ethiopia to be "patriotic" these so-called Patriots never choose local products over international ones even if the quality is equal to the international brand some are going to argue that that's a status symbol maybe it is but if you're not supporting local projects then you're not a patriot go watch this video and boost these local Heroes
vt.tiktok.com/ZSqVdC61m/
instagram.com/reel/DdUbWErPf…
Great job @chapimenge3 very proud of you
21
1,490
What the heck my claude Max usage limit hit without even doing a proper result?
Is Codex Max give more usage limit?
2
5
569
መልካም አዲስ ዓመት 🌼
Meskerem 1, 2019. Ethiopia starts a new year today.
Wishing you peace, health and a year worth the work.
11
230
That's not a recording of me. It's my voice, cloned from one sentence of Amharic.
We're keeping access limited. Consent-only, delete anytime. A public voice cloner puts words in people's mouths, and here that isn't hypothetical.
t.me/dataset_words_bot
7
3
52
1,291
Chapi Menge retweeted
New from Previous Week on Launch.et
Five local tech products emerging with new solutions in the Ethiopian digital ecosystem.
• Harif Moments — A platform designed to capture and share event memories, photos, and videos digitally in high quality.
Visit the product: launch.et/product/harif-mome…
• Dataset.et — An open-source project accelerating local AI growth by gathering voice, text, and translation datasets for local languages.
Visit the product: launch.et/product/dataset-et
• GitStreak — An app that turns developers' daily GitHub coding activity into a motivating habit and challenge.
Visit the product: launch.et/product/gitstreak
• Eagle Pathway — An EdTech platform guiding students from academic success to global scholarship opportunities.
Visit the product: launch.et/product/eagle-path…
• Mezgeb — A minimalist ERP inside Telegram that helps small businesses and individuals easily manage their personal and business finances.
Visit the product: launch.et/product/mezgeb
👉 Explore all products and support the makers with claps: launch.et
#LaunchET #MadeInEthiopia #EthiopiaTech #ProductLaunch
4
11
570
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 👇
4
19
82
2,715
Demo: huggingface.co/spaces/Chapim…
Benchmark, model, code, and all 12 things that went wrong:
huggingface.co/datasets/snap…
Full write-up:
dataset.et/blog/which-amhari…
Free. Built by @snapwre.
4
20
498