PhD | AI & Voice | Startups https://nitter.cf/t.co/PxzriWj2jt
Paris, Île-de-France, FRA
Joined April 2013
- Tweets13.4K
- Following730
- Followers3.8K
- Likes26.7K
Josh Meyer retweeted
A TV playing in the background dropped a voice agent’s completed calls from 20% to 1% in a @veris_ai test.
Machines don’t hear the way humans do.
AI agents now answer more customer support calls, yet they miss what a human agent catches without thinking: a TV, a coworker nearby, a bad line. When that happens, the failure shows up as a bad transcript or an expensive handoff to a human, and teams rarely blame the audio.
Our new post explains what audio intelligence means for CX and contact center leaders, and how Tyto helps your team catch audio failures before callers do:
🔗 ai-coustics.com/blog/audio-i…
Meet Think-Live LLM and Think-Fast LLM, two new LLMs built for voice agents.
Think-Live runs the conversation: when to listen, wait, or respond.
Think-Fast does the work: rules, tools, spoken replies.
Our CPO @edanuff breaks down this, and more from #DeepgramSpeak, in his latest blog: deepgram.com/blog/voice-agen…
We built Polymarket for Voice AI.
Will ElevenLabs v5 drop in the next two months? Will Deepgram Nova-4? Will any transcriber finally get under 12 errors per 100 words?
Place your predictions on markets.speko.ai
Josh Meyer retweeted
another benchmark Phonic-1 tops!
there's a new #1 voice agent on VAmoS Bench 🥇
we benchmarked @phonic_co's Phonic-1 on ultra-realistic voice simulations from @veris_ai
TLDR;
-> 57.5% pass rate (next best: ~45%)
-> $0.20 per call (middle of the pack)
-> strong in street / cafe noise; TV still hard
🧵/
Josh Meyer retweeted
Agentic speech to speech models are not easy to train. Congrats to @phonic_co for leading the VAmoS bench. We benchmarked this model pre release and it’s a strong model.
there's a new #1 voice agent on VAmoS Bench 🥇
we benchmarked @phonic_co's Phonic-1 on ultra-realistic voice simulations from @veris_ai
TLDR;
-> 57.5% pass rate (next best: ~45%)
-> $0.20 per call (middle of the pack)
-> strong in street / cafe noise; TV still hard
🧵/
Josh Meyer retweeted
let's go!! another benchmark where Phonic-1 is the most intelligent while also pareto on latency and cost
there's a new #1 voice agent on VAmoS Bench 🥇
we benchmarked @phonic_co's Phonic-1 on ultra-realistic voice simulations from @veris_ai
TLDR;
-> 57.5% pass rate (next best: ~45%)
-> $0.20 per call (middle of the pack)
-> strong in street / cafe noise; TV still hard
🧵/
huge jump in the pareto curve from @phonic_co ... but VAmoS Bench is still nowhere near saturated 💪
there's a new #1 voice agent on VAmoS Bench 🥇
we benchmarked @phonic_co's Phonic-1 on ultra-realistic voice simulations from @veris_ai
TLDR;
-> 57.5% pass rate (next best: ~45%)
-> $0.20 per call (middle of the pack)
-> strong in street / cafe noise; TV still hard
🧵/
there's a new #1 voice agent on VAmoS Bench 🥇
we benchmarked @phonic_co's Phonic-1 on ultra-realistic voice simulations from @veris_ai
TLDR;
-> 57.5% pass rate (next best: ~45%)
-> $0.20 per call (middle of the pack)
-> strong in street / cafe noise; TV still hard
🧵/
a deeper dive is coming soon. what do you want us to dig into?
-> where it still fails
-> angry vs calm callers
-> noise
-> latency and interruptions
-> something else?
reply and let us know 👇
🧵/
voice AI is definitely having it's moment
two great teams coming together! this is going to have a big positive impact on the voice ecosystem!
congrats both to @deepgramscott and @kwindla 👏
👏👏👏
introducing prism, the world’s first speech-to-speech model you can’t tell from a human in conversation
prism has:
> 300ms response time, 4x faster than OpenAI GPT-Live
> outperforms every other s2s model on intelligence (with NO reasoning)
> the first s2s model customizable through reinforcement learning
we’re a team of 35 with $35M in funding, and we’re unlocking s2s model adoption faster than openai, google, and x.
the first 100 people to sign up here get 1 month free: rime.ai/prism?utm_campaign=p…
Josh Meyer retweeted
Machine Learns - 80 🤖
• Clover: block-diffusion coding LLM
• KronLSTM: linear attention with a convergence proof
• RubatoCodec: adaptive speech tokens
• Speculative Forcing
• MAI-Base-2 pre-training report
• Normalizing Flow Maps
• kechup (mine)
erogol.substack.com/p/machin…
🔥
NVIDIA’s best speech model to date: Nemotron 3.5 Transcribe! It’s an offline model with:
🤖SpeechLLM backed by Nemotron 3.5 Lightning
🧠Steered, promptable, contextual transcription
🌏40 languages
🗣️Up to 8 speaker transcription
⏰Timestamps for each word and speaker
🔥over 2000 RTFx for NVFP4 + DFlash2 on Blackwell GPUs
🏆#1 open model on Open ASR, tied #1 overall on AA-WER, #1 open on David AI multispeaker ASR
🔓Weights, code, recipes, agentic skills, Granary-v2 - all in the open!
See the clip from Her Girl Friday (1940): captioned entirely with a single forward pass!