Federated finetuning of Whisper on Raspberry Pi 5
flower.devI don't think the article mentions it, how well does the rpi 4 and 5 do for inference with whisper especially v3?
v3 only comes in one flavor: large.
I don’t think you’re going to have a good time running the large model on a Pi of any kind.
The large models are 32x slower than the tiny models, roughly.[0]
I just tested, and whisper.cpp on my Pi 4 can transcribe the 30-second a13.wav sample (“make samples” to fetch it) in 18.5 seconds.
You can do the math… 32x = 10 minutes transcribe 30 seconds of audio with the large model. Not a good time for most people.
The Pi 5 could be 2x to 3x faster.
[0]: https://github.com/openai/whisper/blob/main/README.md#availa...
I can confirm that we're seeing 2x to 3x faster (RPi 4 vs RPi 5) in some of our early tests
yes. Finetuning a whisper model on a RPi 5 is ~2x faster than on the RPi 4. Other stages involving data pre-processing with HF dataset is again 2x-3x faster.
One of the Flower maintainers here, we're planning to follow up with a more in-depth performance comparison soon
I’m also interested in peoples’ experience. I’d expect decent performance: Whisper 3 has many model sizes, down to 35Mb, iirc. Training, and especially inference, should be doable on a Pi5.
> Whisper 3 has many model sizes
Nitpick but important - Whisper v2 and v3 are large only. It's actually the same Whisper but the version of the model (large-v2, large-v3) has been updated.
All of the other model sizes are the original release.
I reread your comment multiple times and still don’t understand the important nitpick. Are you saying that the smaller models haven’t been updated alongside the Whisper 3 release? That makes the most sense to me, but I don’t want to misinterpret what you mean!
They only released the "large" model of both v2 and v3, the tiny model is v1
Yes. The example uses Whisper-tiny which is 39M, a perfect match for the downstream task of keyword spotting. Just one line needs to be changed in the code to run a larger Whisper model :)
How would this actually work in practice? Do I ask the user to utter specific words then train on that? How is it different from the traditional speech recognition that I need to 'train' to work better on my voice?
The Holy Grail would be to train the model while using it, without any friction. I don't think these methods support that though.
One of the Flower maintainers here. The code example is primarily meant as a demonstrator to show that it's possible to fine-tune these models in a federated way on devices as small as a Raspberry Pi 5.
The bigger takeaway is that we're close to being able to train/fine-tune models with much better performance by accessing vastly more data on the edge, in a federated way.
The device on the edge creates the data but must also label it, right?
If the task requires labels yes. Alternatively, an auxiliary model (not part of the training) can generate pseudo labels and use those during training. A more general approach (for which a few works in FL have proven to work pretty well) would do instead un/semi-supervised training.
This is cool. This might be a silly question, but what are the scenarios where it's useful for fine-tuning on the edge with small devices? I get inference on the edge, and curious about metrics on that for Whisper, but isn't it better to fine-tune on beefier infrastructure and then deploy it for inference on the edge?
The big opportunity on the edge is access to more data. Especially with the rise of end-to-end encryption, applications will be able to use more (and more diverse) data on the edge to get better model performance. It's generally true that training on beefier infrastructure is easier, but in the long run, nothing can beat access to better data. And edge hardware has gotten a lot faster over the last few years.
It seems like one benefit of fine tuning on the edge is the data doesn't need to move around as much. My father taught me "don't move a pile of dirt twice", so maybe it is like that.
Imagine fine tuning a personal LORA on the end users data. No privacy headaches but all the personalization.
I’m guessing this will also help with thick accents?
yeah. with FL it should be possible to make sense out of all data that is distributed across devices without ever having to move it to a central location (i.e. collect it). In the case of speech data, users participating in a federated setting would likely come from different backgrounds, which could be reflected in their accent or use of language.