Deep learning speech research @GoogleAI. Let's build The Young Lady's Illustrated Primer! Spare time: classical music, science journal clubs, MTB, dance.
San Francisco, CA
Joined July 2007
- Tweets128
- Following478
- Followers678
- Likes320
SF & Oakland: a few friends and I wrote a voting tool as a side project. See 65+ voter guides in one place at openballot.app and go vote on June 2nd! Primaries get less attention than general elections, but in California they're often the election that matters most.
Make voting less painful with Openballot.app - now live for San Francisco & Oakland! Just in time for the June 2026 special elections. Please share with your communities - we’d love everyone to experience how easy voting should be!
daisy stanton retweeted
My team in GDM Frontier AI is hiring (Mountain View). If you're a researcher interested full duplex modeling, multimodal LLMs (Gemini), modality gap, joint speech/text modeling, PGMs, streamable generative models, and representation learning for language modeling -- DM me!
ALT Long-form Speech Generation with Spoken Language Models. ICML 2025 Oral. Recent work from our team on long-form speech language models with state-space models.
daisy stanton retweeted
Wow, google open source some tts models github.com/google/sequence-l…
We also provide reference implementations for the encoder and decoder on GitHub. (Yes, I know no one uses TensorFlow anymore ;P)
github.com/google/sequence-l…
Eric is the powerhouse that got RNN-based TTS models to generalize to utterance lengths not seen in training via Dynamic Convolution Attention (DCA). Now he's at it again with Transformer-based models, using interpolated relative position biases in a novel alignment mechanism 💪
Transformer-based TTS models sound great but have all kinds of reliability issues.
Our new model, Very Attentive Tacotron (VAT), is a Transformer-based TTS system that doesn't drop or repeat words and can generalize to any practical utterance length.
arxiv.org/abs/2410.22179
daisy stanton retweeted
Want a more bike-friendly SF? 🚲 Check out @sfbike's voter guide on OpenBallot along with 40+ others to help shape the city's future! 🗳️🌉
openballot.app/guides/9cba2a…
daisy stanton retweeted
We're thrilled to open sign-ups for OpenBallot in time for the November San Francisco election! See 40+ voter guides in one place, and use our tool to fill out your ballot. You can keep it private or share with friends:
openballot.app/guides
I'm hosting Astral Codex X's San Francisco ballot party on October 19. Come join our engaging discussion: lu.ma/0ufpys63
Very excited to share a side project that we're launching in time for the November election: openballot.app/
daisy stanton retweeted
We're excited to announce the @ClimateChangeAI Summer School 2024!☀️
Are you an #AI expert who wants to tackle #ClimateChange? Are you a #climate expert trying to use #ML in your work? Are you curious about the topic?
Apply & register now! Learn more👉climatechange.ai/events/summ…
DALLE-3 release reminds me of a captivating dinner party art piece I've wanted to build: in-home ASR device -> bit of prompt finessing -> parallel text streams to DALLE-X -> 1-best displayed on a 77" 4k QLED. Result is artistic rendering of conversation on ~30s refresh.
If you haven't listened to Lex's June 30 interview with George Hotz (and especially if you think you hate Hotz), do. It was very good and very entertaining. I recommend audio over video: lexfridman.com/george-hotz-3…
Apparently 1 in ~12 men are colorblind, since red/green color blindness is an X-linked recessive trait. In women, one of the X chromosomes is *randomly* inactivated early in development. But then why are only 1 in 200 women colorblind?
(Also, en.wikipedia.org/wiki/X-inac… is a joyous read)
New work from my fantastic colleagues Soroosh and Matt (they even share an umbrella)!
Very excited to announce this paper, with my incredible co-first-author, @MattShannon123!:
arxiv.org/abs/2212.03232
An attempt to revive generative (joint) modeling as a principled approach to un/semi-supervised learning in seq2seq problems.
Google Research is excited to announce ... TACOSPAWN! It's a TTS model that can speak in human voices that don't exist: google.github.io/tacotron/pu… We learn a prior over the speaker embedding space of a many-speaker Tacotron: arxiv.org/abs/2111.05095
New research from my brilliant colleagues! At long last, direct-to-waveform Tacotron -- using multi-scale normalizing flows (read @rustyryan's thread if you have FOMO and want to learn what those words mean). Everyone's still smiling and kicking ass through pajamafromhome. 🚀
New work with Ron Weiss, @EricBattenberg, @sorooshooryad, @dpkingma -- finally achieving what @log_pie and I set out to do in 2016 before switching to spectrograms: direct waveform generation from characters. (1/7)
abs: arxiv.org/abs/2011.03568
samples: google.github.io/tacotron/pu…
If you train an end-to-end text-to-speech model whose attention fails to align, you can get hilarious descents into speech madness like this: google.github.io/tacotron/pu…. Sadly, @ericbattenberg has reduced the frequency that we'll get to cry with laughter in our office over these...
Check out our new paper on location-relative attention mechanisms for Tacotron TTS. We compare existing mechanism and introduce Dynamic Convolution Attention (DCA), which enables Tacotron to generalize to very long utterances. arxiv.org/abs/1910.10288
daisy stanton retweeted
Detecting deepfakes is one of the most important challenges ahead of us. Following our release of a synthetic audio dataset in Jan, we're releasing a large dataset of visual deepfakes to support researchers working on synthetic video detection #GoogleAI
ai.googleblog.com/2019/09/co…