asitis (@rifeash) header
asitis (@rifeash) profile picture

Interested in speech, language systems and cryptography | Working at @AlpenLabs | Building @__ampixa__

Oblivious RAM
Joined December 2021
Photo by asitis (@rifeash)Photo by asitis (@rifeash)Photo by asitis (@rifeash)Photo by asitis (@rifeash)Photo by asitis (@rifeash)Photo by asitis (@rifeash)Photo by asitis (@rifeash)Photo by asitis (@rifeash)
Khatra TTS is start of ambitious project If this gets selected, I might finally have the confidence to pitch one more proposal
KhatraTTS is our internal FlowLM based text to speech system This will support more than 100 languages and major languages of Nepal Read more: ampixa.com/blog/khatratts-a-… This is still a research preview without any weights released and there are still a lot of questions to answer like - can we separate language from voice when a language has only one speaker - support for low resource language - intonation transfer and such
1
20
665
asitis retweeted
Duty-Free Bits received a Distinguished Paper Award at ACM CCS. The paper makes a versatile cryptographic primitive much more efficient, paving the way for even better Bitcoin bridges. Congrats to @nakstar123 and co-authors Anwesh Bhattacharya and David Heath.
4
8
1
23
1,416
We have something better coming soon Built on top of our new codec. you can check out the development candidate on huggingface.co/ampixa/jyoti
sanoTTS is #1 on Hugging Face’s Open TTS Leaderboard for CPU streaming. 🥇 Our 294K-parameter heart-nano voice starts speaking in 2.8 ms, and three of the top five models are ours. Built in Nepal, measured on the same benchmark as everyone else. Thank you to everyone who uses, forks and builds on sanoTTS. 🤗 huggingface.co/ampixa/sanoTT… 🤗huggingface.co/spaces/hf-aud…
2
23
621
asitis retweeted
why pay for dedicated sim when u can just run ur own ios sim on a crappy linux box. alpha releasing next week!!
Coding agents write mobile code faster than any team can review it. 🙌 Introducing Expo Simulators: cloud simulators built for agents. Your agent runs its change on a cloud simulator or emulator and brings back the recording before it asks for your attention.
2
8
1
34
1,367
Kinda wild to think that services/companies spending money on heedeo are actually saving money
you should kill your CAPTCHAs. for decades we built captchas to keep bots out. but now the bots are your customers. people send AI agents to sign up, buy, and use your product for them. block them and you block real business. let them all in and one person runs 10,000 free trials. captchas ask "are you human?" the right question is "are you a good actor?" so we built heedeo: let good agents in, keep bad ones out. one free trial per person. not ten thousand. heedeo.com :)
1
6
396
Newari is really low resource language. And we managed to build something usable while staying within the constraints Nepali <> Newa voice assistant What's impressive is 33M params streaming ASR and then our flowLM based custom codec based TTS We are doing some good research... but there is much more room for improvement Still on preview. Send me a DM if you want to try
We built a voice assistant for Nepali ⇄ Newari (नेपाल भाषा). Speak one language, hear the other about 2 seconds later. Speech recognition, translation and voice all run in Mac Studio. No cloud API. Read the blog here: ampixa.com/blog/newa-a-nepal…
11
6
1
66
2,542
@nirajbhusal @BeaumontLuke please help reach more people
1
179
When finetuning a TTS model we encounter hindi base most of the times.... Because Nepali and Hindi are really close we can start with Hindi TTS and later use Hindi Leak scorer as a post training signal so Hindi doesn't leak.. Atleast that is the goal Currently in research preview
Our latest work Hindi Leak Scorer for Nepali language If you fine-tune a Hindi TTS model for Nepali and some Hindi pronunciation can stay behind: a vowel shifts, or goes missing. Hindi leak scorer will find the words that sound Hindi for you and you have a training signal Currently in Research preview.
8
780
asitis retweeted
do grounded work before groundbreaking
2
8
498
Ratna is research preview of 32million param Zipformer model built from scratch It's so small that it runs right on your browser on web assembly build... Try now at : ampixa.com/ratna Proof of work: ampixa.com/blog/ratna-small-…
Meet Ratna (रत्न): streaming Nepali speech recognition in a 35 MB file. - Smallest nepali ASR - works right on your browser without API (cloud) or GPU - streaming which means it writes as you speak Try it on your browser right now: ampixa.com/ratna
15
481
This Kukur Tihar we were serious dog researchers... more to come
I only make serious software
3
5
1
77
4,439
The community like to celebrate... and scrutinizing the engineering contribution is expensive. This is becoming embarrassingly cheap and wilddd Why do the hard work of building something genuinely new when you can take an already successful model, fine-tune it, optimize for a handful of benchmarks, give it a new name, and announce a breakthrough?
印度光明节,号称 built in 印度的 Alpie 9B 发布。看了一下,完全是向 Qwen3.5-9B “致敬”: - 架构方面,完全 copy - 分词器方面,词表一个字没动 - 权重结构和数值:基本逐位相同 哦,唯一不同的是模型的自我认知: chat template 每轮都注入指令,问来自哪里就答 “India”。 把 X 网友当傻子么? huggingface.co/169Pi/Alpie_s…
2
6
253
There is a subliminal message when you convert AI companies logos into sound... Look closely at the Mel Spectrogram Turn on your speakers before you start playing
OpenAI CEO Sam Altman believes people owning RAM is wasteful, and that it belongs in data centers instead. "From a global allocation perspective, actually, it’s quite terrible to have everyone have a bunch of the world’s limited supply of DRAM sitting in their own computer."
5
13
1
247
28,280