We turned EmbeddingGemma 2 into a speech-to-text model

371 points by amrrs 11 hours ago on hackernews | 82 comments

Pretty cool application of an embedding model

What about the quality of the transcription?

They say its word error rate is about 5%, so not great. Interesting as a proof of concept. The model this is based on is REALLY tiny.

: koolba | 9 hours ago
Too bad the errors for things like this are not truly independent.
Otherwise could run two passes to get down to .25% or three passes to .0125%. Or four passes to .000625%!
That’d be .02 seconds of error per hour!

robotswantdata | 10 hours ago

WER is not good enough for professional use, fine for hobby projects

mikemzz1 | 10 hours ago

interesting approach. But training ctc head on top of gemma's audio encoder will most likely produce better results and faster inference

: Shershah1024 | 8 hours ago
That's close to what this is: the reader is a CTC head on the audio encoder, but its output words are EmbeddingGemma's own text embeddings for ~229k words, instead of a separately learned vocabulary. I also tried a plain letter-level CTC head on the encoder. It did a bit better on clean audiobooks (7.0 vs 8.2 WER at the time) but much worse on harder audio like a lecture (16.5 vs 10.8), mostly from misspellings.

alberth | 10 hours ago

As someone who knows nothing on how this works, does this mean they have to pay the audio at 200x playback speed to transcribe this fast.

Or is there another approach?

treyd | 10 hours ago

It's not being played through a desktop audio stack, they're operating on the waveform decoded from the audio file. There's no constraints relative to the "natural speed" of the source, just like for video transcoding, it's limited by compute.

radlad | 9 hours ago

Yes but... Frame rate matters with video, which is arguably the video's speed.

: treyd | 6 hours ago
You can flip through the pages of a book much faster than it'd take you to actually read them normally. But if you're putting the book in a scanner then you're still extracting information, faster than the "natural speed" of reading them.

cheschire | 10 hours ago

Probably significant parallelism. Split the audio into a lot of slightly overlapping portions to ensure the transitions between words are caught. I’m sure there’s some relatively simple ways to ensure you don’t double up on words.

drob518 | 10 hours ago

Yea, I was wondering how they dealt with word boundaries and different word audio lengths generally (short and long). Your point about parallelism is interesting. Doing what they say they are doing, this should be ridiculously parallel, possibly with a reduce step at the end to remove redundant words if their slicing algorithm allows for that to happen. But yea, extremely parallel.

: Shershah1024 | 4 hours ago
To answer, words are not extracted in isolation. The model sees a context of about 20 words around it always.

sorou | 10 hours ago

during transcription, nothing is actually "playing" the file in the way a human would to listen to it. the audio is stored as a sequence of bytes representing the amplitude of the physical sound waves over time (the recording device measures, or "samples", the amount that its microphone is vibrating and records that amount, some tens of thousands of times per second) with some metadata and compression applied so that other software knows how to interpret the bytes correctly. this can be given to an audio program to convert the bytes back to physical sound waves using your speakers, but it's just bytes at the end of the day, so any software that can read bytes out of the file can do so at whatever speed it wants, regardless of how long the actual recording is.

so you have a file containing bytes representing the sound wave, and you have existing recordings (read: sound waves represented as sequences of bytes) of people speaking in different languages and contests, and you can feed all those existing recordings into a model along with corresponding transcriptions, and you can get a model out which can identify patterns in recordings of people speaking to then go backwards and see a pattern of sound waves and correctly associate it with a specific token, no need to play anything back. pauses, stutters, etc. are themselves patterns of sound waves that can be identified and worked around to identify just the meaningful parts of speech.

in principle, you could probably learn to do this too, by sitting down for hundreds of hours meticulously starting at waveforms, and this is, very roughly, what signal analysts do: look at waveforms and try to identify patterns. ai just leverages this principle to turn mundane speech into text.

satvikpendem | 9 hours ago

No more than an AI media generator creating the media in realtime.

Imustaskforhelp | 10 hours ago

This has come at just the right time.

Does this work for Hindi or other non English languages though?

I wanted to recently transcribe Hindi + English (mixed) Audio and I used this model[0] and created a custom transcribe command which can give text over input of mp4 files

This took me roughly around an hour or more or around this time frame for a 90 minute transcription source. I would consider myself to be slightly bored during this phase.

I would really appreciate it if something like the project could be made native for a lot more languages but I think that the approach of embedding models and using them in such a way is pretty crazy. Kudos for it and I will try it out!

I can think of many such use cases where Speech to text of even large audio files becomes near instant. As I think someone on hackernews once quoted whom I am very paraphrasing

An order of one or two magnitude of an already existing thing stops becoming a quantity upgrade but rather becomes a quality upgrade in it of itself as it provides access to workflows which previously wouldn't have been feasible.

[0]: https://huggingface.co/parismitaglobalsolutions/indicconform...

tinykerneldev | 9 hours ago

TrueScribe supports tranlating Hindi transcripts in both directions. Not free though.

1970-01-01 | 9 hours ago

I've been using Whisper with great results; GPU acceleration gets the job done, nothing leaves my box, and the medium model is overkill.

https://openai.com/index/whisper/

https://github.com/openai/whisper

: Imustaskforhelp | 9 hours ago
Yeah I will have to try whisper as well. I didn't try it because I wanted to tinker more and also I had thought that whisper was an older model but I think that if it gets the job, then it gets the job done and I will be trying it out on the text that I mentioned, thanks for mentioning it as I will give it another go :-D

fdb | 10 hours ago

Is it open source? They mention something about the Apache License but I can't seem to find a link to GitHub or something.

: cstever | 10 hours ago
I only see that EmbeddingGemma 2 is Apache 2.
"EmbeddingGemma 2 is made by Google DeepMind and released under Apache 2.0. This is an independent project, not affiliated with or endorsed by Google."
: walrus01 | 10 hours ago
https://huggingface.co/google/embeddinggemma-2
: ghboo0927 | 9 hours ago
The reference to Apache 2.0 on the page refers to Google's base model, not this app. Based on the page alone, the app is free, but the source code is not open-source.

jdw64 | 10 hours ago

It would be nice if Korean also worked. I’d like to experiment with a Korean model too.

llimllib | 10 hours ago

Apple's on-device transcription model is about 3x slower than their reported results on my more powerful computer, with a similar error rate in my test (which I consider pretty good results)

I'll test theirs too if I get around to it

Here's a sample CLI wrapping the apple model and providing a simple benchmark: https://github.com/llimllib/transcribe

update: doesn't seem like they provide their model and I don't feel like trying to dig it out of their app bundle, so I'm not going to test this one

s3p | 9 hours ago

Is it even worth using apple's? Maybe they have improved it but when they first started using it to transcribe voice memos it was laughably bad (in my experience). I defaulted to whisper.cpp

llimllib | 9 hours ago

apple's is 5x faster than `large-v3-turbo` with whisper.cpp and in my test had somewhat worse performance, so it's a tradeoff

ChrisMarshallNY | 8 hours ago

I suspect Apple's strengths are in multi-language support.

I use their on-device translation in my apps. Works pretty well. ChatGPT's context-aware translation is better, but it's not on-device.

solarkraft | 5 hours ago

I used it a lot in my AI client before I finally got around to hacking a Deepgram API server into Handy: It sucks enough to give me a headache worrying about its unreliability in every given transcription.

They supposedly improved it on later Apple Intelligence enabled devices, but I don’t have one.

: llimllib | 2 hours ago
my tests are against the Apple Intelligence model, fwiw, not any earlier model

bigtex | 9 hours ago

With MacWhisper you can select the Apple models, it took almost 4x as long to transcribe the same file versus this software and MacWhisper/Parakeet3.

: llimllib | 8 hours ago
interesting! I only tested against Parakeet v2 because my agent suggested it would be faster for English transcription (v3 is multilingual but somewhat slower, it suggested)
On my bench, parakeet-tdt-0.6b-v2 is 2x slower
I wonder if you're using different Apple models when you go through that app? I don't know

vladimirralev | 10 hours ago

Any info on the param size and quantization of each part and in total?

lukko | 10 hours ago

Interesting... is it overkill to use a multimodal model like this on just audio? Could the model size be decreased dramatically, or does the multimodality help the accuracy and speed in some way?

coldtea | 10 hours ago

Correctly transcribing requires context too, no?

: ranger_danger | 9 hours ago
I suppose it depends on the words in question... language knowledge and general world knowledge may not always both be necessary to transcribe accurately.

sd9 | 9 hours ago

Multimodality seems to just help with everything

pjc50 | 7 hours ago

The benefit is that you can take input in one "mode" and extract output in another; that is, the input is audio mode and the output is text mode. The intermediate representation is the "embedding".

spiderfarmer | 9 hours ago

So Marco Arment can sell some more of his Mac Minis.

: Aboutplants | 9 hours ago
First thing I thought of lol

amelius | 9 hours ago

Sounds great if you could use it as a server. Otherwise, why would you want this?

bawana | 9 hours ago

I wonder if this approach can be used to transcribe the soounds made by dolphins and whales

bigtex | 9 hours ago

I tested a 33 minute local audio only file on a M4 Mac Mini with 24GB of Ram and it took 6.75 seconds to transcribe it using this software. Using MacWhisper/Parakeet3 it was about 7.5 secs using a stopwatch.

yaanfpv | 9 hours ago

I think this seems to target English only I believe. Pretty good but thinking os using some of these techniques on our multi-lingual ASR models !

: Shershah1024 | 8 hours ago
Only English for now. I dont see any roadblocks for at least the major European Languages(Not sure about German though with their compound words)

Shershah1024 | 9 hours ago

Hi all. I am Shahir. This was my experiment. First of all - Thanks to whoever posted this on HN. It came as a surprise.

Just to clarify - This started as just an experiment to see if we can use the fact that text and audio share a vector space in Embedding Gemma 2 to convert audio to text. Had to hack around a few things, but, it worked. Speed was just a side benefit. Was not aiming for it.

The accuracy is good, but not in the range of cloud models, but decent enough for everyday use, and this is just a couple of days work. The WER can improve a lot with more training on varied data. Will try that next.

I will opensource the whole thing soon. Might need a few days though. Also, releasing the TTS version for this soon.

Happy to answer any questions you have.

prodigycorp | 8 hours ago

can you turn this into a decision model

drakythe | 8 hours ago

What would a decision model add to and audio transcriber?

: cheschire | 7 hours ago
Reducing cost for a large scale corporation call center. Press or say one.
Basic ML is probably even cheaper. But LLMs might have higher accuracy. “Uhh one I think” = 1
: e12e | 6 hours ago
Theoretically - if the change was deeper than just restricting output - you could feed it thousands worth of audio tokens and get a structured response in milliseconds rather than minutes?
Say classify recorded interviews into viable/not viable candidate?

sureglymop | 7 hours ago

Isn't EmbeddingGemma2 already a decision model? Am I misunderstanding your question?

Shershah1024 | 4 hours ago

It can already classify decently well, and I tried training it as a decision model, It worked well on domains I trained it for. It was beating Liquid's small decision model, but somehow, it never generalized. Out of the trained domains, it was just above chance.

cabarltos | 8 hours ago

Thanks for putting this out there!

xnx | 8 hours ago

Cool experiment. It will be amazing if this can match the WER of the best open weight models at better speed.

People aren't paying enough attention to how great EmbeddingGemma 2 and Gemma 4 are.

: Shershah1024 | 4 hours ago
I am experimenting with the reader with more varied audio. Plus, possibly an ultra fast post processing layer.

usern20260720 | 7 hours ago

can this tool transcribe latin language texts? McWhisper allegedly does it but it outputs gibberish

rahimnathwani | 6 hours ago

It was your post yesterday that made me aware of granite-speech-5.0-470m-turboctc (which I guess you used as a comparator because it's already very fast).

Thank you!

  uv tool install mlx-audio
  mlx_audio.stt.generate \
  --model ibm-granite/granite-speech-5.0-470m-turboctc \
  --audio.wav \
  --output-path transcript

(Until now I've been using mlx-community/parakeet-tdt-0.6b-v3)

: Shershah1024 | 4 hours ago
Yes. Granite is really fast. IBM has actually been releasing some great models that are flying under the radar

samuelbalogh | 9 hours ago

Fantastic! Any plans to support other languages?

Shershah1024 | 8 hours ago

Yes. Planning to train a multilingual reader. Have not done a prototype yet.Will do one once I have the right dataset

: samuelbalogh | 8 hours ago
Awesome work, thank you!

saaaaaam | 8 hours ago

I recently swapped to using Parakeet v3 on my old M1 Macbook Air, using MacWhisper. It was so fast I thought it had failed. As in, by the time I clicked "transcribe" and looked at my screen again a few seconds later the transcript had finished. Absolutely insanely fast.

: 1-6 | 3 hours ago
People say old MacBook Air M1 but its architecture was so groundbreaking that it's still relevant in 2026. Apple definitely underplayed its release in 2020.

sbdaman | 8 hours ago

Maybe it's just a COVID time blending together effect but it still blows my mind how quickly we went from "automatic audio transcriptions are not even worth having on" to "an hour of audio transcribed in 13s locally."

Pre-pandemic it seemed like even the very best in class auto-transcription features were just unusably bad. I guess that was like 7 years years ago though.

: gambiting | 8 hours ago
I'm just waiting for someone to tell the YouTube team. Hopefully 10 more years and someone over there will realize it's a thing.

whstl | 7 hours ago

For people interested in super quick models:

"One cool trick" I like for running those kinds of models very quickly on macOS is converting them to CoreML so they run in the Neural Engine.

I had amazing success with Spleeter, Demucs, MVSEP-MDX23-music-separation, DeepFilterNet, Whisper, Hunyuan3D for 3D mesh generation. Plus a lot of others for both work and leisure.

I also convert the code to Swift without dependencies, but that's a personal preference.

The runtime on some of them went sometimes from 60s to 10s. And since it's just Swift, I can just give them to non-technical co-workers and they don't have to setup a whole Python environment.

If you don't wanna do this by hand, this is something that existing LLM models are good at. Even Luna can do it without much fuss.

bbor | 7 hours ago

Ok let's talk about the really important stuff here: that model name is atrociously long for such a game-changing model, and "embedding" is already an awkward phrase. I propose... Embedda?

: Shershah1024 | 4 hours ago
He he. Thanks for the suggestion. Just wanted to show that an embedding model can transcribe. Maybe we should call it emscribe :D

motoboi | 7 hours ago

A nice side effect is that it transcribes in English an audio in Portuguese. And it actually make sense.

But the overall transcription quality is very low.

Silagi | 6 hours ago

Is it low quality in the "It didn't match the meaning that was conveyed" sense, or in the "It didn't capture the exact words used" sense?

My reaction to seeing the project was that being an embedding model it'd be pretty good at conveying the meaning but not great as using the same terms, is that what you're seeing?

mahlerbox | 5 hours ago

Curious to hear the answer to this as well. Presumably it would be stronger on sentences with a high proportion of cognates/English terms with Latin roots? Does the English output over-represent words of Latin origin?

: Shershah1024 | 4 hours ago
The reason it generalizes to english now is because the reader is trained to identify English words, and it obviously understands Portuguese, so it just gave the text in English. Once I train it on multiligual vocabulary, it should be able to handle Portuguese

cabarltos | 6 hours ago

You should add in a more comprehensive Audio Input selection - for example, selecting "Input 3" on my audio interface microphone

drakirus | 5 hours ago

I'd be curious to see the WER on out of vocabulary words, seems like this cannot generalize to unseen word's given this step: 'The closest of ~229,000 English words wins, and the slices become a transcript'

: Shershah1024 | 4 hours ago
You can just embed those new words and they get added to vocabulary. It is fairly simple. One risk is - When a new word is very close to a known one, it tends to fall to it.

guigoguedes | 5 hours ago

Does it work for portuguese?

mrkn1 | 5 hours ago

yapsnap is faster on CPU and uses a zipformer https://github.com/kouhxp/yapsnap/

p0w3n3d | 5 hours ago

The downloaded model is stored in

  Library/Application\ Support/ai.haku.eg2demo

raajg | 4 hours ago

This approach is interesting. Does this mean that I can limit the the words to the ones I actually use and build a tiny model that finds the closest match to only those words I use in 99% of my conversations?

armcat | 3 hours ago

I applaud you on the experiment, this is a great showcase in how the latent space is shared between modalities and a great performance achieved.

You stated WER of 5% - against which benchmark?

Whistle [1] for example is a 16MB STT model but WER varies between 4-26% depending on the benchmark.

[1] https://cactuscompute.com/blog/whistle

andai | 55 minutes ago

Nice. Would love to see a comparison with Parakeet and the recent Needle. Are there any benchmarks for transcription speed and accuracy?