3 2 4

Mayowa Daniel

mayowadan

AI & ML interests

None yet

Recent Activity

new activity 29 days ago

deepseek-ai/DeepSeek-OCR-2:Made a Demo for people who want to quickly try the model

liked a Space 3 months ago

ResembleAI/Chatterbox-Multilingual-TTS

new activity 5 months ago

ibm-granite/granite-docling-258M:Hallucinations

View all activity

Organizations

None yet

New activity in deepseek-ai/DeepSeek-OCR-2 29 days ago

Made a Demo for people who want to quickly try the model

#2 opened about 1 month ago by

ysong21

liked a Space 3 months ago

Chatterbox-Multilingual-TTS

🌎

366

Chatterbox TTS supporting 23 languages

New activity in ibm-granite/granite-docling-258M 5 months ago

Hallucinations

#39 opened 5 months ago by

ntsarb

Why is granite-docling-258M so slow?

🤝 1

#37 opened 5 months ago by

hgarp-prozis

liked a model 6 months ago

calcuis/chatterbox-gguf

Text-to-Speech • 0.3B • Updated Sep 18, 2025 • 1.76k • 51

replied to Xenova's post 6 months ago

Amazing work!
Where can I find the unminified code for the demo?

reacted to hexgrad's post with ➕ 8 months ago

Post

4181

IMHO, being able & willing to defeat CAPTCHA, hCaptcha, or any other reasoning puzzle is a must-have for any Web-Browsing / Computer-Using Agent (WB/CUA).

I realize it subverts the purpose of CAPTCHA, but I do not think you can claim to be building AGI/agents without smoothly passing humanity checks. It would be like getting in a self-driving car that requires human intervention over speed bumps. Claiming AGI or even "somewhat powerful AI" seems hollow if you are halted by a mere CAPTCHA.

I imagine OpenAI's Operator is *able* but *not willing* to defeat CAPTCHA. Like their non-profit status, I expect that policy to evolve over time—and if not, rival agent-builders will attack that opening to offer a better product.

2 replies

reacted to hexgrad's post with 👍 8 months ago

Post

6197

I wrote an article about G2P: https://hf.co/blog/hexgrad/g2p

G2P is an underrated piece of small TTS models, like offensive linemen who do a bunch of work and get no credit.

Instead of relying on explicit G2P, larger speech models implicitly learn this task by eating many thousands of hours of audio data. They often use a 500M+ parameter LLM at the front to predict latent audio tokens over a learned codebook, then decode these tokens into audio.

Kokoro instead relies on G2P preprocessing, is 82M parameters, and thus needs less audio to learn. Because of this, we can cherrypick high fidelity audio for training data, and deliver solid speech for those voices. In turn, this excellent audio quality & lack of background noise helps explain why Kokoro is very competitive in single-voice TTS Arenas.