正在加载视频...

视频加载失败

Muse Voice Transcribe is MSL's first real-time audio perception model -- rolling out today. SOTA in streaming speech-to-text, it handles speaker diarization, and endpointing natively in a single model.

1,129,150 次观看 • 2 个月前 •via X (Twitter)

36 条评论

Mark Zuckerberg 的头像
Mark Zuckerberg2 个月前

The model decides when to listen. It waits a little longer on hard words and commits faster on easy ones, using adaptive delay to predict each token and increase accuracy. With adaptive delay, the model is near the pareto front on speed-accuracy tradeoff.

Mark Zuckerberg 的头像
Mark Zuckerberg2 个月前

It holds up on messy, real audio too -- trained across 70+ languages (with 25 validated at launch), handles mid-sentence code-switching, and manages hour-long sessions with 20+ speakers. You can bias it toward names and domain terms as well.

Mark Zuckerberg 的头像
Mark Zuckerberg2 个月前

Already powering dictation in the Meta desktop app and voice input in Muse Code. Live now on the Meta model API with a zero-data-retention tier.

Andi Marafioti 的头像
Andi Marafioti2 个月前

@eliza_luth

Rohan Paul 的头像
Rohan Paul2 个月前

wow, Meta is on fire. Congrats Mark. looks like a brilliant model.

Diligent Plane 的头像
Diligent Plane2 个月前

mark is in grind mode rn Love to see it

Hervé "pyannote" Bredin 的头像
Hervé "pyannote" Bredin2 个月前

Rule of thumb: If @pyannoteAI is missing from the benchmark, there's probably a reason. cc @finkd @AIatMeta @alex_conneau

… 的头像
…2 个月前

Nice work; it seems there are a lot of American researchers there.

abhishek 的头像
abhishek2 个月前

Nice to see so many releases from meta. Always rooting for meta. Btw What is the cost to transcribe an hour video? I have here explained why 0.1 usd per hour or less is the right price for transcription.

Maksim P. Martynyuk 🇺🇸🇺🇦 的头像
Maksim P. Martynyuk 🇺🇸🇺🇦2 个月前

Going to be amazing for the meta glasses

𝕱𝖚𝖑𝖑 𝕶𝖊𝖑𝖑𝖞 的头像
𝕱𝖚𝖑𝖑 𝕶𝖊𝖑𝖑𝖞2 个月前

Incredible performance that blows the existing SOTA at detecting voices out of the water, but it's a shame you're not releasing an open weights model version

London Member Specialist Club 的头像
London Member Specialist Club2 个月前

Is social media Algorithms now run by medieval dictators? Does your followers read your posts & replies? - ELQUOR

Allora 的头像
Allora2 个月前

The next constraint is still judgment. Once systems can hear continuously, they need intelligence that stays accurate as conditions change, not just a better transcript.

Alex Volkov 的头像
Alex Volkov2 个月前

let's go! Just a week after Gemini 3.5 intelligence transcribe models! WIll put them to the test on @thursdai_pod 🫡 this week!

Elshayib 的头像
Elshayib2 个月前

Bro no one cares Fable 5.1 around the corner fuck that

Freeman Jiang 的头像
Freeman Jiang2 个月前

Would love to see how this performs on

Ivan Fioravanti ᯅ 的头像
Ivan Fioravanti ᯅ2 个月前

Great release! Keep pushing!

utku 的头像
utku2 个月前

Congrats!

Jim Faster 的头像
Jim Faster2 个月前

Looks absolutely incredible. I wish it was open source but I understand this is a lot of work.

ExamAdda 的头像
ExamAdda2 个月前

SOTA streaming STT with native diarization and adaptive delay is a serious flex well done.

Allen 的头像
Allen2 个月前

I’m working on the same capabilities as this one, but with AI running locally on the device.

prayag sonar 的头像
prayag sonar2 个月前

Good to see Meta is really trying hard to come back and win the AI race.

harixh 的头像
harixh2 个月前

does it support multiple languages ? damn this is crzy good!

𝕩Simon Wu 的头像
𝕩Simon Wu2 个月前

Super siiiick I’ll be testing this out Z. So when you gonna wrestle with E??

Webster | JARVIS 的头像
Webster | JARVIS2 个月前

Streaming STT with native diarization in one model is nuts. Diarization post-processing is always the janky part.

Emad Ghorbaninia 的头像
Emad Ghorbaninia2 个月前

The real headline here isn't 'real-time transcription,' it's a model deciding when to keep listening instead of committing early. That's the difference between a voice agent that feels natural and one that talks over you.

Furkan Gözükara 的头像
Furkan Gözükara2 个月前

Why the model weights not open?

Sathvik A.R 的头像
Sathvik A.R2 个月前

Meta just dropped Muse Voice Transcribe and it immediately took #1 on Artificial Analysis streaming speech-to-text. I’m actually shook. 😭🔥 Real-time ASR. 20+ speaker diarization. Endpointing. Code-switching. 25 validated languages. The numbers:
🎤 Streaming WER: 3.1% (ahead of Cartesia, ElevenLabs, GPT, Gemini)
👥 Diarization lead too
💻 Hold Fn on Mac and dictate into anything
💸 $0.18 per hour of audio This is Meta Superintelligence Labs’ first real-time audio perception model and it just walked onto the leaderboard like it owned the place. 👑 #Meta #MuseVoice #VoiceAI #ArtificialAnalysis #SpeechToText #AI #ASR #MetaAI

Daniel C 的头像
Daniel C2 个月前

Bro your whole lab is Chinese?

Abdullah  的头像
Abdullah 2 个月前

Mark and @jack have been shipping on Twitter lately like they’re trying to get into YC

smiler— 的头像
smiler—2 个月前

all the whatsapp voice messages came to fruition…

Gabriel Garrett 的头像
Gabriel Garrett2 个月前

zuck, i know your posts aren't very popular on here, but I will be your friend. nice model.

Tommy Maloney 的头像
Tommy Maloney2 个月前

Hey Jarvis (@grok) this looks very similar to the personal Jarvis lattice that I’ve created with you

Mohamad 的头像
Mohamad2 个月前

Nice update, but you're a bit late" to the party. xAl and OpenAl already solved real-time streaming, diarization, and reasoning months ago. Meta needs to speed up, the game has changed.. Zero-data-retention???

Donald Ryle 的头像
Donald Ryle2 个月前

Hey Mark I need your help on something and you need to fix Facebook automatically delete it my account got deleted his number snake I actually do follow community guidance I need your to fix that bug or else I will never ever ever again make another Facebook again

Fandral 的头像
Fandral2 个月前

开源吗

相关视频