Alex

@fluidinference , Audio AI & LLM ML Engineer

[Fluid Inference @fluidinference]

NVIDIA's Nemotron 3.5 ASR Streaming Multilingual is now available through FluidAudio optimized for Apple Silicon so apps can run ~40-language real-time ASR entirely on device, no cloud required.

Apps shipping with it today include
@SpokenlyApp, @ALTIC_DEV, @SnaplyAI, and others.

FluidAudio:
https://t.co/VWaAPfLRh5

Original model:
https://t.co/87wRTlO5dm

@NVIDIAAIDev

[Iago Cavalcante @iagoangelimc]

LeafTok narrates books on-device using FluidAudio for TTS.

Zero network calls during playback. Voices ship with the app.

The trade-off vs cloud TTS: bigger app, fully offline.

[Hamza Qayyum @hamza_q_]

Impressive. Very nice.
Recently contributed a patch for Parakeet V2/V3 to FluidAudio that makes it match this speed, on-device, i.e. ~300x speed factor. Can transcribe an hour of audio in ~12 seconds on an iPhone. And tdt-ctc-110m is ~33% faster than that (1 hr in ~9s).
https://t.co/zcRQmvxO1f

[Spokenly @SpokenlyApp]

Spokenly now supports NVIDIA Nemotron 3.5 ASR Streaming Multilingual via @fluidinference — 40 locales, on-device, free, no subscription. Privacy-first dictation, now global.

App: https://t.co/l3j1rxlTMA
Model: https://t.co/FVfly8ZW8a
@NVIDIAAIDev #VoiceAI

[ALTIC @ALTIC_DEV]

FluidVoice 1.5.15 now supports @NVIDIAAI Nemotron 3.5 ASR Multilingual collaborating with @fluidinference - 40 language-locales, free forever, fully on-device, and ultra fast!
👉Download here —> https://t.co/lprygBMhxg

@NVIDIAAIDev @NVIDIAAI #NemotronSpeech #VoiceAI

[Giacomo Venier @giacomovenier]

Snaply now supports NVIDIA Nemotron 3.5 ASR Multilingual via @FluidInference — 40 language-locales,
on-device, no cloud, free to use.

@NVIDIAAIDev @NVIDIAAI
#NemotronSpeech #VoiceAl

[Barathwaj Anandan @BarathAnandan7]

Today's a big day for Nemotron models.

Along with Ultra, we also shipped Nemotron Speech 3.5 that now supports 40 Languages and it's insanely Fast and Ultra Low latency!

I collaborated with @Alex_tra_memory, @fluidinference and @ALTIC_DEV to port the model to coreML to make bring the latest Nemotron model to any macbook using FluidVoice!

Give it a try and lmk what you think!

Link below ⬇️

[Sachin Desai @sach1n]

Nemotron ASR Multilingual running on an iPhone 17 Pro in CoreML.

Many thanks to @fluidinference for the CoreML model and to @NVIDIAAI @NVIDIAAIDev for the model itself.

[Sachin Desai @sach1n]

Supertonic3 running on an iPhone 17 Pro using ANE on CoreML. It’s blazing fast with low RAM consumption and background capable. 2 mins worth of audio generated in 3 secs.

Many thanks to @fluidinference for the port.

[Audivize @audivizelabs]

Audivize now supports NVIDIA Nemotron 3.5 ASR Multilingual via @fluidinference, adding support for 40 language-locales all on-device.

Demo: https://t.co/RwWiEsPBe6

Model: https://t.co/aBzlrWGWkO

@NVIDIAAIDev @NVIDIAAI
#NemotronSpeech #VoiceAl

[Alex @Alex_tra_memory]

when you write with coding agents you are also feeding the company training data. reducing the skills and ppl needed for complex or huge projects.

[Alex @Alex_tra_memory]

@hamza_q_ haha yeah, i think what happened was these coding agents collected enough training data from their users. the agents don't struggle as much with the more complicated OS environment

[Alex @Alex_tra_memory]

gave claude on windows another try, they have def improved the agent to operate much better on windows now. lesss environmental and system troubles, much more ease of operations

[Alex @Alex_tra_memory]

@hubeiqiao @Joe_Speaking let's gooooooo

[Alex @Alex_tra_memory]

just finished chapter 1 of wdndev's llm_interview_note book. i had asked LLM to help quiz me on the concepts. the first chapter is impressive and does well in showing us why transformers were needed. i don't think i even understood half of the content. onto chapter 2 the main piece.
i am already starting to doubt if i am can even become a ML engineer

https://t.co/XLPEPJMamG

[Alex @Alex_tra_memory]

this is like the best resource i have ever read about LLMs, the og is chinese but with deepseek you can easily translate or ask it to explain to you.
its fluid clarity
https://t.co/RMxAVsQG4I

[Alex @Alex_tra_memory]

i. do not have a math background yet alot of finetuning models for ANE optimization is intuition maxing thanks to coding agents.

[Alex @Alex_tra_memory]

alot of metalearning is related to understanding whats the fundamental unit a certain field use to measure or benchmark "progress". like how you can't have an economy without an agreed upon currency

[Alex @Alex_tra_memory]

tried to fuse mel spectrograms and encoder and decoder models for Parakeet RNNT and CTC before btu the enigmatic ANE Scheduler doesn't make this easy

[Alex @Alex_tra_memory]

Weight matrices and Upsampling is the deduction of additional information from limited information. i.e 6*6 matrix input into a 8*8 upsample matrix based on the position and values of the individual coordinates