Alec Freudenstein retweeted
When we were smaller, we used to live and breathe off of free credits
Paying it forward!
Alec Freudenstein retweeted
Introducing Exa Websets V2
Build massive datasets from the web, now 4x faster and with 2x lower cost per result.
Also new:
- 20% higher accuracy and recall
- Pay-as-you-go pricing, no subscription
- API access for all
The Hume Audio Decision API in 30 seconds, for anyone who missed the launch.
Ask it anything about a recording and it answers from the sound itself, not the transcript. A sneeze, a shift in tone, a change of accent. Answers in under a second.
One model for every question you have about audio.
Replying to @ElevenLabs
@ElevenLabs v4 Turbo and v4 just joined the Real World VoiceEQ leaderboard at 3rd and 4th of 38.
We ran both through 18 of our human evals, three generations per prompt, with native-speaker raters for non-English passages.
Alec Freudenstein retweeted
Super excited to be hosting the second SF voice research club.
Paper presentations by @nishnik on Meta Muse TTS and Taejin Park on Nemotron 3 Diarization.
Here is a link to the event if you would like to join!
luma.com/yd93ae7y
Alec Freudenstein retweeted
Even if AI were conscious, how could we possibly know which tokens felt good or bad to it?
Text-to-speech was built to read a single voice aloud. Now it's being asked to hold a conversation between two, and that's a different problem entirely.
The speakers have to stay distinct, trade turns with natural timing, and keep their emotional states from bleeding into each other. None of that shows up in word error rate.
So we built a rubric: five dimensions, 48 two-speaker scripts, and four scenarios designed to break a dialogue by pairing similar-sounding voices, forcing frequent interruptions, contrasting emotions, and running long exchanges.
Three blind human raters per clip. We ran three generations of Google's Gemini TTS models through it, alongside other multi-speaker systems.
Keeping two voices apart is the hardest part of the job. Every system handles a male and female pair easily and drops on at least one same-gender pairing. The newest Flash model made its biggest gain on exactly that.
Models with the same overall score get there differently. One leads on keeping speakers’ voices distinct, another on making the dialogue sound like one continuous recording. The overall score hides that difference.
New capabilities need new evaluations.
Full case study, with audio examples, is linked in the comments.
Alec Freudenstein retweeted
Proud of the team for delivering the most natural and expressive speech model I’ve ever used (by far).
One simple insight was that text-to-speech isn’t a “small model” problem anymore - the more frontier the model, the better it understands what it’s saying, the more it will sound like it understands what it’s saying.
This isn’t just text-to-speech, but frontier audio generation, including an endpoint to design and edit voices using text prompts. Try generating voices with any nuanced dialect, accent, or generational speaking style - AI should communicate in the exact style that specific users are most comfortable with. Or fun things like “Vampire meditation guide with Transylvanian accent,” “Swedish ASMR furniture review,” “goblin auctioneer selling rare mythical objects,” “tinny monotone robot selling neural transplant upgrades,” etc. AI doesn’t need to sound human, it should just be maximally understandable (and fun to use!).
Audio is the primordial way humans communicate. As we approach human-level speech capabilities - which has admittedly taken a bit longer than text - I expect it to become the primary way we interact with AI, online in real-time applications and offline in education, entertainment, etc. And everyone should have their own trusted AI they recognize by voice.
We’re launching Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS ⚡️
Our most expressive audio models yet let you create custom voices across 100+ languages or pick from 2,000+ ready-to-use ones. You can direct back-and-forth conversations, guide the delivery line-by-line, and add natural cues like <laughs> or an active listening interjection like |mhm| all while generating hours of consistent, glitch-free audio.
Sounds pretty cool, right? So… how should you use them?
— Gemini 3.8 Flash TTS: Need to design bespoke vocal personas from scratch and with line-by-line level control? This is the model! Built for high-fidelity creative production like gaming, immersive audiobooks, and podcasts.
— Gemini 3.8 Flash-Lite TTS: Want the AI to automatically adjust its tone and pacing on the fly for near real-time voice agents? This is your engine! Built for cost-efficient scale, high-volume dubbing, and bulk audio creation.
Alec Freudenstein retweeted
Treble's voice simulation platform is used by voice AI model developers, AI wearable, and robotics companies
spr.ly/6011BGpPYB
Alec Freudenstein retweeted
Two years at DoorDash. I do not say this lightly.
We are extremely close to the burrito arriving before you decide you want it.
We are not asking for a ban. We are asking for a pause.
Alec Freudenstein retweeted
Just increased the ChatGPT Voice usage limits:
Go: 3h Live mini
Plus: 3h Live
Pro 100: 15h Live
Pro 200: Unlimited
More details: help.openai.com/en/articles/…
Alec Freudenstein retweeted
Introducing Muse Voice Transcribe, the first real-time audio perception model from Meta Superintelligence Labs.
Muse Voice Transcribe delivers real-time streaming ASR, diarization with 20+ speakers, and endpointing. It’s multilingual with seamless code-switching and improves accuracy with language, keyword, and context biasing.
The model ranks first on @ArtificialAnlys streaming speech-to-text and on public diarization benchmarks.
For anyone who saved the paper and hasn't opened it :) huggingface.co/blog/asr-benc…
We silenced a number in the benchmark audio. It was never spoken.
Some of the highest-scoring speech recognition models wrote it down anyway, in 30–40% of cases.
New research with @huggingface on measuring when a model is optimizing for the test instead of the audio.
Alec Freudenstein retweeted
Introducing GEN-1.5, a one-shot learner.
It can learn new tasks in a few seconds. Show it what to do, and it generalizes.
This capability emerged from pretraining on physical data at scale, as a step towards our mission of building general intelligence for the physical world.
Alec Freudenstein retweeted
Still looking for 1x more intern at Salesforce AI Research! If you have experience or are interested in designing benchmarks/evals, I’d love to hear from you! 🚀
In addition, we have a couple more research scientists/engineering spots open! Have any experience with voice models/agents, reach out ASAP 🚨
Alec Freudenstein retweeted
Running this experiment has been very fun and informative about how good LLM systems can be
This Post is from a suspended account. Learn more
Alec Freudenstein retweeted
Since January, @SierraPlatform has grown revenue >100%, while growing our engineering team <15%.
Small, AI-native teams are more productive, cohesive, and fun. Here are some examples of how we work 🧵
Alec Freudenstein retweeted
it's another great day to tap the sign
Lilian Weng, the Thinking Machines cofounder who announced her departure earlier this week, citing startup-related stress and illness, is rejoining OpenAI.
She'll be working on using AI to develop new models (recursive self-improvement).
w/ @amir
theinformation.com/briefings…



























