Nvidia Nemotron 3.5 Transcribe: open ASR for 40 languages | Hacker News Reader

Nvidia Nemotron 3.5 Transcribe: open ASR for 40 languages

huggingface.co

134 pointswondersky26 comments

ASR = automatic speech recognition. I think this is the same as STT, speech to text?

It seems to want an NVidia GPU. I didn't look closely to find out if it can run usably without one. For me, usable might mean 10% of realtime, i.e. an overnight run on my old x86 box for 1 hour of transcription. Of course faster is always nice. YMMV.

Almost all ML models come with reference python code that will run on most architectures.

If you want an implementation optimized for a certain architecture the cost is between 10-30USD of opus tokens and is oneshottable: ("implement Nvidia nemotron 3.5 transcribe for the intel npu on this machine. Use the hf transformers implementation as reference. The finished artifact takes pcm on stdin and utf8 chunks on stdout".)

You can substitute your old intel CPU by modifying the prompt by replacing "intel npu" with "the target cou on this machine ... Your task is complete when the model decode performance is 80% of the memory read speed of this machine"

Thanks, that's interesting, I hoped maybe it would use pytorch or something. But what I'm wondering is whether the result would run at usable speed. I don't currently use anything like opus and only want to use open weight models, and at the moment I don't have a GPU machine, though I guess I can spin up an hourly VM with one sometime.

Parakeet does really, really well in CPU only. I would try it first.

> ASR = automatic speech recognition. I think this is the same as STT, speech to text?

I'm surprised no one corrected you. There's no speech left when it's converted to text, while, as you stated, the other one is Automatic speech recognition :-)

> Transcribe is built for recorded audio, not live streaming.

Bit unfortunate! Nemotron Streaming in Handy has basically solved dictation for me. It recognizes German mixed with English terms and brand names really well.

Can you share some info how you got that working?

https://handy.computer/

Select nemotron in the model selection

Oh, since it's called "handy" I expected it would work on mobile phones, as per the name suggests. That seems it only has desktop operating systems.

Also, there's no mention of Nemotron in that handy. Is that the same Handy GGP was talking about?

There are apps on the Google Play Store where you can run on-device models or connect to OpenRouter where they are run for you. Im using Qwen 3 ASR locally on my android, it's the second most accurate open model after the one in this post.

what app do you run on Android for this use-case?

https://github.com/DevEmperor/DictateKeyboard. I forked it to add more models (with Claude). I use openrouter models as well as the local ones

https://handy.computer has Nemotron Streaming and it works fabulously well on any Mac down to the base M1. I also built a little Deepgram compatible API into it that I use with a phone app.

From the model card, this supports near-field only speaker transcription, experimental but still something not available in any other OSS Speech LLM.

I played around with Nemotron 3 recently. It's OK, but makes an error every few sentences, at least on the data I worked with (standard vocab). I'm sure we'll get there, but I was pleasantly surprised by how I can recognize words better than it can.

Agreed. I find Parakeet or Qwen ASR better.

I’m still sad Voxtral 3 wasn’t made open source by Mistral.

Yes if parakeet would work then thats the answer and the code has already been written.

Otherwise the answer is really mechanical. It's will run as fast as you can get the weights into your cou's ALUs. since it's 30b-a4b, if you have 30G spare ram you can load the whole thing and it will run about as "fast" as any 4b model (qwen3.5-4b, gemma4 e4b etc).

Otherwise you will be streaming from disk/SSD which may or may not be very slow.

kinda makes sense how Nvidia gets more distribution after acquiring hugging face, right?

i believe OSS STT is amazing but htings break udring mulaw and turn detection where closed source models still are ahead

Voxtral 3 is probably the better product, but this is cheap!

How is it for Japanese

testing it against parakeet right now for ASR, it's slower, heavier, and has a higher error rate. but i'm doing a pass with the embedded translation to see if that's any good, it might be worth it.

e: also i'm testing it with a Q4 quant that might not be full accuracy

ugh, translation is worse than passing a parakeet transcript to a cloud llm. disappointing model.

tried running it against a full anime episode, both vocal isolated and raw. it got confused by english words in the theme song and failed spectacularly three times

I have a few thousand audio interviews as MP3, around an hour each and with specialized vocabulary. Typically there are no more than 2 to 3 speakers. Audio quality ranges from mediocre to fair to good.

I was thinking of creating a local app that calls OpenAI’s VTT model. In my most recent evaluation, the OpenAI model was best for my use case. But is that indeed the case?

We may never know.

Reply on news.ycombinator.com