Alec Freudenstein
389 posts
Alec Freudenstein
@AlecInRealTime
Speech Datasets & Evals @ Hume AI || Knicks in 5 💙🧡
New York, NY Katılım Ağustos 2021
727 Takip Edilen224 Takipçiler
That's our founding engineer at Willow and he's using @WillowVoiceAI :)
swedishasian67@michellezfr
Wearing noice cancelling masks to talk to Claude is crazy
English
Today, we’re announcing Bonsai 27B: the first 27B-class model to run on a phone.
Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops.
Until now, models in this class have been impractical to deploy locally. A 27B model occupies roughly 54 GB in 16-bit precision, and even a strong 4-bit build is around 18GB - too large for a phone and for most laptops.
Bonsai 27B changes that.
It comes in two variants:
• Ternary Bonsai 27B: 5.9 GB, 1.71 effective bits per weight, optimized for laptop-class quality.
• 1-bit Bonsai 27B: 3.9 GB, 1.125 effective bits per weight, optimized for phone-class footprint.
Everything is open-sourced today under the Apache 2.0 license.
English
You probably think you know what your customers are calling about. One travel company did too. Then the data said: laundry.
emerj.com/why-agentic-ai…
#Dialpad #AICX #customerexperience @comcast @Emerj
English
@ai_coustics Love it! Congrats to the ai-coustics team! Y'all are moving at warp speed right now!
English
We asked Sasha from our team to introduce our latest model in the chaos of the real-world and she never got to finish a sentence.
That's kind of the point.
English
We extend trust to a voice before we've even heard what they are saying. We do it with each other all the time. A measured tone reads as competence. Warmth reads as care. Good pacing reads as "this is under control."
The same holds true when the voice is automated. A frustrated customer decides how a call will go in the first few seconds, and the main signal they use is the sound that greets them. Most CX automation gets this backward: teams obsess over what the system says and treat how it sounds as decoration. In a tense moment, the sound is the message.
There's science behind it. In voice-to-voice conversation, people automatically reproduce the emotional tone of whoever they're talking to. The effect is fast, unconscious, hard to suppress. Put plainly, calm is contagious.
That's what we built Lexis for.
Read more 👉 eu1.hubs.ly/H0wXXbF0
English
One of the most rewarding parts of open research is watching the community take your work further than you could on your own. It's been just 6 months since we launched 🤏Tiny Aya 🤏 and here's a look at all the exciting projects coming out of Expedition Tiny Aya:
Cohere Labs@Cohere_Labs
Tiny Aya is proof that multilingual AI doesn't have to be massive or limited to a handful of languages. In recent months, researchers worldwide have used Tiny Aya to build projects in education, accessibility, safety, interpretability, & more. 🚀🤏🌎 cohere.com/blog/tiny-aya-…
English
English
Tomorrow's medicines should be designed with the precision and scale of modern engineering.
I'm thrilled that @ChaiDiscovery has raised a $400M Series C at a $3.8B valuation from @IndexVentures @KleinerPerkins @Sequoia @_DimensionCap to accelerate progress towards that goal.
Chai Discovery@chaidiscovery
We’ve raised $400M at a $3.8B valuation to further advance AI-driven molecular design. The round was led by @IndexVentures, alongside @KleinerPerkins, @Sequoia, @_DimensionCap, and others.
English
Nikita deserves his flowers. There’s been a HUGE improvement on my X feed recently.
The FYP has been…dare I say…full of content that I actually find useful or relevant to my work.
Nikita Bier@nikitabier
We're rolling out a small tweak to boost visibility of your posts to your mutuals (people who you follow back). We noticed this data was missing from the algo and it made your friends appear less in your replies. This resulted in the reply section feeling more like a battleground with people you don't recognize. This should also help clusters form around interests more easily, which many people have asked for.
English
Alec Freudenstein retweetledi
An initial startup idea can't usually be both grand and precise. In practice they're usually either grand and vague or precise and small. Precise and small is better. You know who your initial users are, and you expand outward. With grand and vague you can't even get started.
English
Alec Freudenstein retweetledi
I’m hiring for my research team @OpenAI 🪄
Data will pave the way for AGI. This is a foundational TPM role to shape frontier AI models with real world data, RL environments, and data acquisition.
Looking for:
- Entrepreneurial, gritty, high horsepower
- Highly technical and deeply curious about AI research
- Excited to lead relationships with CEOs and data partners from day 1
English
Inference-time reasoning/scaling is also useful for TTS, especially for offline or latency-insensitive use cases.
When improving the model’s ceiling becomes difficult, we can still improve final output quality by sampling multiple candidates and selecting the best with a reward model.
For example, generate audio with different temperatures / sampling settings. The same TTS model can produce outputs with noticeably different quality, speaker similarity, prosody, and style. Then use best-of-N selection to choose the strongest candidate.
The trade-off is simple:
No model weight update is needed, but inference cost increases by roughly N times. This makes it less suitable for real-time TTS, but very attractive for offline generation, dubbing, content creation, or high-quality batch synthesis.
Two useful reward-model patterns:
ORM: Output Reward Model
Score the complete generated audio after it is finished.
This works well for metrics that require the full utterance, such as ASR-based WER/CER.
PRM: Process Reward Model
Score partial generations during the generation process.
This is useful when the reward can be computed on chunks or segments, such as speaker similarity. The reward can guide which token/audio path to continue.
My takeaway:
For TTS, inference-time reasoning/scaling can act like a “quality booster” on top of a fixed model.
ORM = best-of-N selection;
PRM = reward-guided generation.
It does not replace model training, but it can push better outputs out of the same checkpoint.
English
English
Voice AI updates don't stop during the summer! ☀️
On the market side, @Gradium secured a $100M funding round led by NVIDIA, alongside news of voice AI agents expanding into points of sale, banking, and clinical use cases. @HydawayDigital introduced real-time enterprise audio deepfake detection, and @mywhispp raised €5 million to scale its real-time, on-device voice reconstruction AI.
On the platform side, @OpenAI is leading the updates this week, confirming the rollout of GPT-Live (its full-duplex, speech-to-speech model) and releasing a faster version of GPT-Realtime. Really impressive! 👏
🪄 On the STT and TTS side, @cartesia introduced Ink-2 for STT. Meanwhile, for TTS, @SpaceXAI released 21 new flagship voices for Grok Voice, and @omilialtd launched Lexis, their ultra-fast model featuring voice cloning for Contact Centers.
Finally, @livekit added a feature to provide feedback during long tool calls and released an interesting update for generating context to improve STT accuracy 🎉
Read the full update with more news and links in the newsletter below 👇
English
"Result: the balance mechanism works, but the click is fragile and the wall is permanent."
okay Claude...
English
Introducing Flash-MSA, the world’s first open source sparse attention training kernels optimized for extreme context lengths. 4 simple kernels, 400%+ speedup over dense flash attention at long context ⚡
English
@eshamanideep @OpenAI capitalism is so beautiful, isn't it? 🥹
English
After having proved Fable 5 and GPT 5.6 Sol, Fable, even with the guardrails, is so so so better ...
English