Alan Cowen

Life update: I’m now at @GoogleDeepMind Feels surreal to not be at @hume_ai anymore, but the team there is in an incredible position to continue advancing cutting edge voice AI research. There are changes coming in how people interact with AI. There are changes coming in how labs will teach AI to make people happy. Voice intelligence is increasingly seen as a key unlock on both fronts, a thesis Hume was very early to adopt. So Hume is sharing the infrastructure for voice data, evals, and RL to help advance voice intelligence and AI empathy across the industry. It’s proving to be a much bigger opportunity than expected. I’m excited for the Hume team and grateful to everyone who helped us get to this point. Will share more on what I’m working on soon

@deedydas Either uses an internal tool that goes by a different name (like everything at Google) or is in a small minority of holdouts

Google Senior Staff Engineer to me: “Yeah, I have no clue what Claude Code / Codex is but I hear it’s all the rage. No, I don’t really care, I just need GOOG to hit $400 and keep this job for 2-3 more years so I can retire!”

Today we're releasing our first open source TTS model, TADA! TADA (Text Audio Dual Alignment) is a speech-language model that generates text and audio in one synchronized stream to reduce token-level hallucinations and improve latency. This means: → Zero content hallucinations across 1,000+ test samples → 5x faster than similar-grade LLM-based TTS → Fits much longer audio: 2,048 tokens cover ~700 seconds with TADA vs. ~70 seconds in conventional systems → Free transcript alongside audio with no added latency

@guitchounts @GoogleDeepMind @hume_ai Yep! That's something Hume is doing with a lot of frontier labs

@AlanCowen @GoogleDeepMind @hume_ai Are there any plans for Hume to share emotion voice data?

Life update: I’m now at @GoogleDeepMind Feels surreal to not be at @hume_ai anymore, but the team there is in an incredible position to continue advancing cutting edge voice AI research. There are changes coming in how people interact with AI. There are changes coming in how labs will teach AI to make people happy. Voice intelligence is increasingly seen as a key unlock on both fronts, a thesis Hume was very early to adopt. So Hume is sharing the infrastructure for voice data, evals, and RL to help advance voice intelligence and AI empathy across the industry. It’s proving to be a much bigger opportunity than expected. I’m excited for the Hume team and grateful to everyone who helped us get to this point. Will share more on what I’m working on soon

@krishnanrohit @GoogleDeepMind @hume_ai Thanks!

@AlanCowen @GoogleDeepMind @hume_ai Congrats !!

@OfficialLoganK @GoogleDeepMind @hume_ai Likewise!

@AlanCowen @GoogleDeepMind @hume_ai Welcome to the team! Excited to work together :)

@Scobleizer @GoogleDeepMind @hume_ai Thanks for all your support Robert. What you do for the AI startup community is so important

@AlanCowen @GoogleDeepMind @hume_ai Congrats! It's going to be interesting to see what you do there. Thanks for coming on my show. In your early days, it was a joy to watch you.

@timekeepur @GoogleDeepMind @hume_ai It will continue and get better

@AlanCowen @GoogleDeepMind @hume_ai Congrats Alan - longtime devs utilizing Hume at DistrictZero 👏🏽- Will API the service continue?

One performance, infinite voices. Voice Conversion is now live on Hume’s creator studio and API! Generate the same pacing, pronunciation, and intonation with one recording across any voice you choose. Hear it for yourself ⬇️

@dflieb Been waiting for this for a while

The AI device I think I want: - AirPods form factor - with wide angle cameras for context - with always-on ChatGPT voice mode - that knows when I’m talking vs other sounds (which it’s terrible at now) - that can respond reactively or be proactive I don’t think I want a display. I’ll have my phone in my pocket anyway.

Excited to be powering @NianticSpatial ‘s Dot

Hume社との提携について、著名なAI研究家である清水亮さん @shi3z に取材していただきました。

@supertradeish @hume_ai It should be possible to reproduce any accent reliably using voice cloning. If you test a specific accent, let us know how it works!

@hume_ai Impressive improvements in speed and cost! How does Octave 2 handle regional accents within those 11+ languages?

Introducing Octave 2: our next-generation multilingual text-to-speech model What’s new: - Fluent in 11+ languages - 40% faster (<200ms latency⁠⁠) & 50% cheaper than Octave 1 - Multi-speaker conversation - More reliable pronunciation - New voice conversion & phoneme editing capabilities For the month of October, we’re offering 50% off our Creator plan - use code OCTAVE2 at checkout!

@heyaytac Thanks for the feedback, we are working on it! It might already be better tomorrow

@AlanCowen Overall great progress but just a small feedback: I have played around a bit with ev4 mini and it lacks a lot of nuances in the german language, especially when pronouncing words correctly. It also switches from german to english when repeating a phone number.

we've been working on making voice AI faster and more realistic. we're hoping that scaling voice will give AI compatibility with the human psyche: every voice session is time-locked, interruptible, and rife with feedback. but first, we need higher quality voice experiences.

@hume_ai 日本語に最適な音声AI

Today @hume_ai is releasing its latest text to audio and audio to text AI models. Remember this company? Brings emotions into your audio in a way others don't. But now better and here founder @AlanCowen gives me a deep dive into the audio/AI space and an update on their latest, coming later today.

@Noahpinion That’s because there’s a third possibility thought to be most likely: AI doesn’t plateau, but training costs shrink by orders of magnitude relative to inference costs (because the scale of deployment grows much faster than the scale of training)

Leading AI labs treat model training expenses as capex. However, if AI continuously improves, training the next model is a cost that never goes away. That's closer to opex. And if AI plateaus, trailing firms catch up and the product becomes a commodity.

@willdepue They mean learning from production data with an objective function that approximates user satisfaction

everyone who mentions ‘continual learning’ as a problem is usually just talking about sample efficiency. clearly, you should ‘continually learn’ by continually training trajectories back into the model! there’s no mystery: this just doesn’t work with low sample efficiency.