asitis

@aabhpsy pretty soon dai. But first OCR... thanks for the product idea

@rifeash Live translate in the original form?

For each answer that gets it right, the answer + person name will be in early supporters section when we release this

@cold_daal @__ampixa__ Thank you so much

@__ampixa__ Congrats on the launch. I remember testing out the tts voice ranking site you guys posted on reddit.

🇳🇵 kala-tts : नेपालमै बनेको, पहिलो आफ्नै देवनागरी G2P सहितको खुला-स्रोत नेपाली VITS आवाज। cloud छैन · तपाईंकै CPU मा चल्छ। pip install kala-tts · 🎧 tts.ampixa.com/kala

@shreemaan_abhi @__ampixa__ Thanks... more interesting things coming. Its been long time working on stealth . aba haluka significant kura yesari nai share gardai janchu hola

@__ampixa__ Great work dai and team. Now I can connect the dots backwards when you shared that tts ko graph draw gareko whiteboard diagram several months ago. Rooting for you!

@aabhpsy @__ampixa__ You can also check Higgs Audio v3 TTS by @boson_ai huggingface.co/bosonai/higgs-…

Any already available options even on GPU? Can we continue work from what you have done so far and use GPUs to get production grade cloned audio in natural Nepali like we speak? I don't understand all but I can dig in and learn from 0 but since you guys have pioneered, it would be nice to get insights.

@bbk_dkl Do you mean the full conversational voice agent? For that we require speech language models ... We might be able to join ASR + language model + TTS and in text to speech part we should be able to use it but i wouldn't think that would be the good usecase of this model here.

@rifeash Great. Can it be used for full duplex model?

Paper + writeup coming soon

any pattern of numbers can be described as a sum of cosine waves .

Opensourcing today

@bijaysenihang I understand you no longer access to fable as a Nepali citizen 😆

I am done staying in Nepal and creating hope, only for the pathetic Nepal government to break you from every side. Going forward, I will no longer provide my knowledge, time, or cybersecurity expertise to the Government of Nepal.

The world i grew up in no longer exists.

Design: model translates Limbu to Nepali, then in a completely fresh context translates its own Nepali back into Limbu. Scored vs the original human Limbu (chrF). An "echo" metric catches models that fake it by copying the input through both legs which disqualified our chinese brethrens

Can frontier LLMs actually translate Limbu which is Kiranti language of eastern Nepal I round-trip-tested them on 100 human-reviewed phrases from Nepali school-curriculum materials What i used: Grade-1 math textbook, translated into Limbu by humans.

@rishadbaniya well it's my sister voice.

@rifeash idk why familiar voice xa

if this post gets 5 likes. i will open source it.

Something big to come within these 6 months. Each day is going to be exciting

i look forward to our chinese brothers liberating the knowledge from within fable-5 and selling it to me at 5% the cost & 2x the speed

exploration will lead to preservation. Exciting things will drop this week

She said "I'm fine", but my speech language model didn't understand her. Because it doesn't catch tones, emotion and stress. Here is how to solve it if you take ASR like whisper and it's 16th decoder layer(P_16). Then create a reconstructor model it is trained on three passes first pass: audio, text -> whisper -> P_16 + text second pass: p_16 + text--> reconstruct mel spectogram third pass: compare with original mel spectogram Train until reconstruction is perfect. Now you can replace that p_16 on whisper_v3 and get a new model. Call it whisper pro Now use the whisper pro +(texts, audio, emotion metadata) from emotion set like (IEMOCAP, CREMA-D) to create a SLM(Speech language model) input = [P₁₆ prosody vector] + [text token embeddings] Congrats you got a better SLM arxiv.org/abs/2605.05927